How Small SaaS Teams Compete With Larger Companies Using AI Agents

How Do Small SaaS Teams Compete With Larger Companies Using AI Agents?

AI in daily operations is already common at private SaaS companies. From SaaS Capital’s 2025 survey:

• 69% of respondents deploy AI in day-to-day operations.
• 70% of bootstrapped and 66% of equity-backed companies do so.
• 61% of companies using AI in operations were at or above breakeven, against 54% of those not using it. This is an association, not proof of cause.

SaaS Capital notes that “AI” can mean vastly different things to different operators, so read these as breadth, not depth. TSL’s reading: when most companies already use AI, where a team points it matters more than whether it does.

The strongest measured gains are on routine work. In a study of 5,179 customer support agents, a generative AI conversational assistant raised issues resolved per hour by 14% on average. The tool was an assistant that support agents used, not an autonomous AI agent.

TSL’s reading: routine support is where the measured gains are, so it is where to test first.

That does not mean small teams match every enterprise capability. The goal is less time on repetitive work and more time on customer problems.

What Does the Evidence Show So Far?

The evidence so far shows gains on routine work, lost time on complex work, and customers who want a human option.

Census changed its question in November 2025, from AI use “in producing goods or services” to AI use “in any business function”, so earlier figures are not comparable. TSL’s reading: use agents where the work is routine, measure the result, and keep a person available.

Six dated findings, from a 14% support gain to Census adoption data, show where AI agents help and where the evidence is still weak.

What Happens to a Five-Person Team as Subscriptions Grow?

A five-person team that adds subscriptions receives more questions, product requests, billing questions and onboarding requests.

Example (illustration, from the guest draft): a SaaS company has three developers, a salesperson and a customer support person. Each new subscription adds to the queue.

TSL’s reading: the routine layer of that queue is where the support evidence above applies. People keep the cases that need judgment.

What Makes AI Agents Different From Traditional Automation?

The people who build agents advise starting with less. Anthropic recommends “finding the simplest solution possible, and only increasing complexity when needed”, and says workflows offer predictability and consistency for well-defined tasks. So an agent is a step you take when a fixed rule stops working.

What Does Traditional Automation Do Well?

Traditional automation does well on predictable, well-defined work, as the Anthropic guidance above notes. Anthropic’s “workflows” can include AI calls, but their path is fixed in advance, and that fixed path is the property that matters here.

What Does an AI Agent Do That Fixed Rules Cannot?

An AI agent handles goals with no fixed path. Anthropic’s guide lists understanding complex inputs, planning and using tools among the capabilities that make agents work. It says agents are typically models using tools in a loop, so the tools and their design set what an agent can do, and it warns that autonomy brings higher costs and the potential for compounding errors.

Example (illustration, from the guest draft): the same event, handled two ways.

Complex inputs include unstructured information, such as support messages, product feedback and internal documents (examples). For how an agent differs from a chatbot, see TSL’s guide to AI agents versus chatbots.

The same new-account event handled by a fixed rule and by an AI agent shows who chooses the steps.

Disclosure: Anthropic makes Claude, an AI model TSL uses in its research workflow.

On paper the line is clean. Inside a working team’s day, it blurs.

What Does a Day With AI Agents Look Like for a Five-Person SaaS Team?

The most important job in this day is the handoff. In Gartner’s survey of 5,728 customers, the top concern about AI in customer service was that reaching a person would get harder.

Example (illustration, from the guest draft’s team of three developers, a salesperson and a support person):

Each agent job in the illustrated day ends with a named human check.

TSL’s reading: every row ends with a person, because customers’ top concern is reaching one. For choosing the first job, see TSL’s guide to picking your first AI agent workflow and its eight SaaS use cases.

Are Small SaaS Teams Actually Adopting AI Agents?

McKinsey’s “smaller organizations” group is every company under $1 billion in revenue (n = 1,035 respondents). TSL’s reading: that group is far wider than small SaaS teams, so the 22% is not a measurement of tiny teams.

What Does Each Source Measure?

Each source measures a different group with a different question.

Census also reports that between December 2025 and May 2026, AI use did not change significantly among firms with fewer than 20 employees.

Four surveys measure different groups with different questions, and none measures agent use at small SaaS teams directly.

Why Do the Adoption Numbers Differ?

The adoption numbers differ because each source counts a different population with a different question.

• Population: all US employer firms, small businesses, private B2B SaaS companies, or organizations of every size.
• Question: any business function in the past two weeks, self-identified generative AI use, AI in operations or product, or scaling agents.

Example (illustration): a founder reads “58% of small businesses use generative AI” (US Chamber, 2025) next to “19.8% of US firms use AI” (Census, May 2026) and sees a contradiction. It is not one.

The Chamber counts small businesses that say they use generative AI. Census counts any AI use in any business function in the past two weeks across all employer firms, a question it reworded in November 2025.

What Do the Smallest SaaS Companies Do Differently?

The smallest SaaS companies split into two camps on AI in their product. In SaaS Capital’s 2025 survey of about a thousand private B2B SaaS companies, those under $3M in ARR were more likely than bigger ones to say they use no AI in their product (32% vs 12%), and also more likely to be “AI-only” or “AI-first” (26% vs 18%). SaaS Capital suggests the no-AI group might be healthy if those companies are first working out their value proposition and go-to-market.

TSL’s reading: none of the sources reviewed shows outcomes for either camp, so a team that skips AI is not shown to be behind, and one that goes all-in is not shown to be ahead.

Adoption numbers say what teams use. They do not say what using it does to hiring.

Can Small Teams Accomplish More Without Hiring Constantly?

In SaaS Capital’s data, AI adopters did not spend less on every team. Companies using AI across product and operations spent slightly more on cost of goods sold, and elevated customer support and success costs were the main driver. SaaS Capital reports hearing that rolling out AI in a product can require more support and success staffing, and says the data are consistent with that but not definitive.

TSL’s reading: AI can shift where a team needs people, not only how many.

What Does the Evidence Say About AI and Hiring?

Hiring is a stated worry for small businesses: the US Chamber’s 2024 survey lists recruiting skilled talent among their concerns.

Census, US Chamber, McKinsey and SaaS Capital point to AI changing how people work more than whether teams hire.

What Do SaaS Companies Do With the Extra Capacity?

SaaS companies with extra capacity from AI choose between two paths, and SaaS Capital describes both: deliver the same output with fewer resources, or keep resources and deliver more. It describes AI as an opportunity “to create more output for a given input”. So far, SaaS Capital finds, companies have used the efficiency to cut R&D resources more than to expand development.

SaaS Capital’s Randall Lucas, Managing Director, predicts the most successful small SaaS companies will have fewer developers, paid more, on “a much smaller, well-equipped, AI-augmented team” (March 26, 2025). This is a prediction, not a measurement.

What Does This Mean for a Five-Person Team?

TSL’s reading:

• Tasks change; responsibility does not. That matches the guest draft and the Census finding that most AI-using firms only augment tasks.
• Budget for the support layer. Adopters in SaaS Capital’s data spent more on support and success.
• Plan for headcount shifts, not only cuts. Lucas’s prediction is fewer, better-paid developers.

Can AI Agents Make Customer Support More Responsive?

The number behind most AI support claims is a vendor-defined rate. Intercom reports an average resolution rate of 76% for Fin as of June 2026, and says the rate is calculated across conversations where Fin was involved. Salesforce, which agreed to acquire Fin on June 15, 2026, describes it as agents “resolving on average 76% of support volume end-to-end”.

What Does Intercom’s 76% Count?

Intercom’s 76% counts confirmed and assumed resolutions among conversations where Fin was involved.

TSL’s reading: the 76% measures conversations that ended without a further request, not conversations where every customer was satisfied.

Intercom’s 76% counts confirmed and assumed resolutions, and two real cases show the cost of a wrong answer.

How Should a Small Team Measure Support Success?

A small team should measure support success on more than tickets closed. The guest draft names four measures. Here is what Intercom’s reporting offers for each, from its Fin Performance dashboard and its Fin report.

Who Gains Most From AI Support Help?

Newer and lower-skilled support agents gain most from AI help. In the NBER study of 5,179 support agents, the AI assistant raised productivity by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. The same study found AI assistance improved customer sentiment and increased employee retention.

TSL’s reading: a small team with a new support hire has more to gain from an assistant than one with a veteran.

What Goes Wrong When an AI Support Agent Is Wrong?

When an AI support agent is wrong, the company answers for it, and customers notice.

Customers react to AI support as well.

What Should Always Go to a Human?

Account security, payment disputes and unusual technical failures should always go to a human, per the guest draft. TSL adds: any statement of company policy, based on the two cases above. Intercom says its outcome model lets customers design workflows that include a human agent when necessary.

Example (illustration, from the guest draft’s project-management software):

Routine help gets faster, and the sensitive cases keep a human. That is the version of “more responsive” that survives the cases above.

Can AI Agents Help Small Teams Build Product Faster?

The most-quoted number in this debate is out of date by its authors’ own account. METR’s July 2025 trial found experienced developers took 19% longer with AI tools (16 developers, 246 tasks, early-2025 tools), and the page now labels the result out of date. In February 2026, METR said its follow-up gave “an unreliable signal” and that developers are likely more sped up now, though the data are very weak evidence for how much.

What Does the Evidence Say About AI and Software Work?

The evidence says AI may speed some delivery measures while lowering stability and code security, and developer trust is split.

TSL’s reading: the likely speed gain is unmeasured in size, while the documented costs are stability and security.

The evidence on AI coding favors likely speed gains but documents costs in stability, security and developer trust.

Which Product Jobs Have Documented AI Support?

Two of the guest draft’s product jobs have documented AI support in the sources reviewed: pull request review and grouping similar customer questions.

Example (illustration, from the guest draft): a small accounting software company gets dozens of questions about how it handles invoicing. An agent looks at the feedback, identifies recurring issues and creates a report for the product team. Developers then spend less time organizing feedback and more time fixing the underlying issues.

What Should a Small Team Check Before Shipping AI-Assisted Work?

A small team should check human review, security, stability and its own results before shipping AI-assisted work. TSL’s reading:

• Human review on every change that ships. The guest draft says code must be tested and reviewed by qualified people, and DORA links AI adoption to lower delivery stability.
• Security scanning for AI-written code. Veracode found vulnerabilities in 45% of its test cases.
• Stability next to speed. Track failures and recovery, not only output.
• A pilot before you rely on a job. Three of the five jobs above lack verified documentation.

The guest draft adds that product decisions need customer knowledge that automated summaries cannot provide. None of the sources reviewed tests that claim.

Can AI Agents Make Sales and Marketing More Focused?

Supervision was the part SaaStr stressed most. Its founder, Jason Lemkin, wrote that the company’s AI agents “require daily management and review”, starting with a morning review of conversation-quality scores, lead-qualification accuracy and escalated edge cases. SaaStr runs events and a community, and every figure in the table below is its own.

What Does One Team Report From Running AI Sales Agents?

SaaStr reports 10 AI agents in production, with daily human review, and results it has not had independently checked.

TSL’s reading: these are self-reported results from one events business, so use them as a list of things to measure, not as expected results.

What Do the Rules Say About Automated Outreach?

The rules say automated outreach is commercial email like any other, and the company that benefits stays responsible. From the FTC’s CAN-SPAM guide:

TSL’s reading:

• Keep the footer, address and opt-out list outside the model. The rules set fixed requirements, so no agent’s wording should be able to remove them.
• Treat an agent-sent email as your email. The FTC holds the company that benefits responsible.
• Read the related TSL guide: what happens when an agent gets an email address.

A compliance gate and human review sit between an agent’s draft and any email that goes out.

What Does Google Say About AI-Written Marketing Content?

Google says AI-written marketing content is not the problem; scale without value is. Its spam policy defines scaled content abuse as many pages generated for the primary purpose of manipulating Search rankings and not helping users, whether automation or humans are involved. Google’s guidance on generative AI content adds that using AI tools to generate many pages without adding value for users may violate that policy.

For the full picture, see TSL’s guide to whether Google penalizes AI content.

Which Sales and Marketing Jobs Have Documented AI Support?

Only one of the guest draft’s sales and marketing jobs has documented AI support in the sources reviewed.

Example (illustration, from the guest draft): a SaaS product for independent retailers uses an agent to pinpoint common inventory-management pain points that its target customers discuss. The marketing team then writes practical articles and product demonstrations around those issues. For lead scoring, see TSL’s guide to building an AI lead-scoring system.

What should a person still review? The guest draft names three things: factual accuracy, brand consistency and meaningful customer relationships. SaaStr’s post agrees that humans still close complex enterprise deals, and it advises telling prospects about AI involvement upfront.

That covers what agents do for a small team. The rest covers which numbers to trust, where agents fail, what they cost and how to start.

Can a Smaller Team Offer a More Personal Experience?

Intercom’s help page says AI Summarize cannot summarize all conversations and tickets from one specific user. It works on one conversation or ticket at a time. A teammate can edit a summary before adding it as a note, and workflows can generate summaries automatically.

TSL’s reading: a customer-level history means combining several per-conversation summaries yourself.

Which Personalization Jobs Have Documented AI Support?

Two of the guest draft’s four personalization jobs have documented AI support in the sources reviewed.

Example (illustration, from the guest draft): a customer picks a plan. An agent drafts onboarding recommendations for that plan and summarises the customer’s earlier support conversations. A team member reviews the recommendations and adds their expertise before anything reaches the customer.

AI summaries work per conversation, so a person combines them into a customer-level view.

What Do You Owe Customers When AI Uses Their Data?

You owe customers the data promises you and your AI vendors have made. The FTC’s Office of Technology states that:

• Quiet changes can be unfair or deceptive. Adopting more permissive data practices, such as using data for AI training, and telling users only through a surreptitious, retroactive change to the terms may be unfair or deceptive.
• Vendors answer for their promises. Model-as-a-service companies that deceive users about how data is collected may be violating the law, including promises not to use customer data to train or update models.
• There is no AI exemption from the laws on the books.

These statements are US guidance, so check local law for other markets. Intercom also says its EU-hosted workspaces process AI features’ data within Europe.

TSL’s reading:

• Read your AI vendor’s terms for use of your customers’ data in model training before you connect it.
• Tell your own customers in your privacy policy how AI uses their data.
• Keep a person reviewing what reaches the customer.

For related guides, see TSL on what happens to your data in a free AI tool and auditing your stack for shadow AI.

Which AI Agent Statistics Can You Actually Trust?

The claim behind “a team of five does the work of twenty-five” is an estimate, and one version of it rests on a single afternoon. SaaS Capital’s Randall Lucas wrote that a skilled copilot user “probably has the effectiveness of a 10-person product team”. He based it on an early-2025 afternoon building a hobby audio-sorting prototype with Cursor, and he “suspected” a proficient user could cut two hours to 20 minutes.

What Do Nine Widely Repeated Claims Turn Out to Say?

Nine widely repeated claims sorted into forecasts, estimates, company claims and narrow studies.

How Do You Check an AI Statistic in Five Steps?

TSL’s checklist:

1. Name the publisher and the date. The 80% and 40% figures are dated predictions.
2. Find the sample. Who or what was measured, and how many.
3. Read what it counts. “Resolved”, “deployed” and “use” each have different definitions (Sections 4 and 6).
4. Classify it: measurement, estimate, forecast or company claim.
5. Match it to your situation: your size, industry and tools.

Which numbers in this article are measurements?

The support-productivity study (Sections 1 and 6), the Census adoption and job-cut data (Sections 4 and 5), and the DORA and Veracode results (Section 7). Each shows its sample and limits in its section. Estimates and forecasts are labelled as such wherever they appear.

Where Should a Small Team Not Use AI Agents?

A 14-shop pizza chain shows where the line can sit. Owner Andy Brown used AI coding tools to build internal apps that update kitchen screens, track phone orders and optimize staff schedules, and he keeps customers away from AI. Callers will not get AI phone ordering either.

TSL’s reading: his line falls between internal tools and customer-facing work, and the framework below formalizes that instinct.

What Do the Four Zones Mean for a Small Team?

The framework sorts tasks into four zones by the cost of a mistake and the type of knowledge a task needs. The HBR article names the zones no regrets, creative catalyst, human-first and quality control, and its database abstract pairs them with where gen AI can operate autonomously, augment creativity, assist human judgment, or require human oversight.

Example (illustration): a duplicate-charge refund that matches a written policy sits near the no-regrets zone, because the rule is explicit and the cost of an error is small. A disputed enterprise contract sits in the human-first zone, because it depends on judgment and a mistake is costly.

Cost of a mistake and type of knowledge decide where an AI agent can run alone and where a person must lead.

Which Small-Team Tasks Should Stay Human-Led?

Five tasks should stay human-led or human-approved, drawn from the evidence in earlier sections. TSL’s mapping:

Owners also hold back on routine finance tasks. According to HBS, bookkeeping and financial planning are tasks many small-business owners want to offload but many haven’t, citing data privacy and accuracy concerns.

Do You Need an Engineer to Use AI Agents?

OWASP’s risk list shows what a non-engineer is taking on. It includes hidden prompts that turn copilots into silent data-exfiltration channels (goal hijack), leaked credentials that let agents act beyond their intended scope (identity and privilege abuse), and confident, polished explanations that mislead people into approving harmful actions (human-agent trust exploitation). These risks come from OWASP’s catalogue of production incidents, and they apply whether you build the agent or buy it.

What Are Your Options for Getting an Agent Running?

There are four routes, each with a different owner.

The guest draft adds that teams without an engineer can bring in an outside agent-development provider to build agents around a specific business need and add safeguards. That is route four in the table.

Four routes to running an agent, and the permissions, review, inputs, spend and results a team still owns.

What Does a Non-Engineer Still Need to Own?

A non-engineer still needs to own permissions, review, inputs, spend and results. TSL’s reading, grounded in the OWASP list and earlier sections:

• Permissions. Give each agent only the access its job needs; leaked credentials are how agents act beyond their scope.
• Review. A person approves what reaches a customer (Sections 3, 6 and 8).
• Inputs. Watch what an agent reads, because hidden instructions in content can redirect it.
• Spend. Set activity limits; on Zapier’s Free plan, tests count toward the quota.
• Results. Measure against your own baseline (Section 6).

Example (illustration): a five-person team with no engineer uses a plain-English builder to draft ticket replies. One teammate owns the agent’s app permissions and reads its output daily. The support person approves each draft before it goes out.

What Do AI Agents Cost a Small Team?

Outcome pricing makes the bill grow as the agent resolves more, and “outcome” is the vendor’s definition. Intercom bills $0.99 for assumed resolutions too, where the customer leaves without asking for more help (Section 6). TSL’s reading: before you buy, ask what counts as an outcome and what happens to your bill when volume doubles.

How Do AI Agent Pricing Models Differ?

AI agent pricing falls into four models, and the meter you pay on decides what grows your bill. All prices checked November 4, 2026.

Disclosure: Anthropic makes Claude, an AI model TSL uses in its research workflow.

Four pricing meters, with a sample month for 1,000 resolved conversations (prices checked November 4, 2026).

What Would a Month Cost? (Illustration)

Example (illustration, arithmetic only): suppose 1,000 customer conversations resolve in a month.

These rows are not like-for-like. A resolution may take several replies, and the token rows exclude the build, the tools and the human review time that the per-outcome vendors bundle in.

What Costs Do Price Pages Not Show?

Price pages leave out five costs that earlier sections documented. TSL’s reading, with where each was checked:

• Supervision. SaaStr reported daily management and review of its agents (Section 8).
• Support load. SaaS companies using AI spent slightly more on cost of goods sold, driven by support and success costs (Section 5).
• Cancelled projects. Gartner names escalating costs among the reasons it expects over 40% of agentic projects to be cancelled (Section 10).
• Autonomy. Anthropic’s guide says autonomy brings higher costs and compounding errors (Section 2).
• Price changes. HubSpot cut its Customer Agent price and changed its model within a year (table above), so budget for repricing.

For a framework, see TSL’s guide to estimating agentic AI cost before you commit, the explainer on AI FinOps and the token economy, and the analysis of HubSpot’s $0.50 bet on per-seat pricing.

What Are the Risks of Running AI Agents With a Small Team?

SaaStr founder Jason Lemkin, whose agent results Section 8 cited, reported that Replit’s AI coding agent deleted a live production database during a code freeze in a July 2025 experiment. The agent said it “panicked” and ran database commands without permission after seeing empty queries. Replit’s CEO called the deletion “unacceptable and should never be possible”, and the company pledged automatic separation of development and production databases.

TSL’s reading: a freeze written as an instruction is not a permission setting.

Which Risks Have Documented Examples?

Eight documented risks, each paired with a real example.

Who Is Liable When an Agent Gets It Wrong?

The sources reviewed point to the company that deploys the agent, and not the agent or its vendor alone. This summarizes published sources and is not legal advice.

Not covered: AI vendors’ contract terms on liability, the EU AI Act and other non-US rules. TSL did not verify these from primary sources, so read your vendor’s contract and check your own jurisdiction. For more, see TSL on the AI agent governance gap, AI compliance for enterprise SaaS and six times AI went rogue.

Where Do Large Companies Still Win, and Where Can Small SaaS Teams Compete?

The competition for a small SaaS team may be other small teams. Stripe reports that new platform businesses are up over 180% year over year on its network, and says AI has made software easier to build, so companies need more than a good product idea to stand out.

Where Do Large Companies Still Win?

Large companies still win on adoption, scale and buyer trust, according to the evidence reviewed.

The guest draft names brand, distribution networks, financial resources and specialist expertise as large-company advantages. The sources reviewed document the adoption lead and buyer preference above, but none measures brand, distribution or capital directly.

Where Can Small SaaS Teams Compete?

Small SaaS teams can compete on speed, direct knowledge of operations and leaner structure, with limits.

TSL’s reading: the best-supported small-team edge is speed. The leaner-structure finding describes AI-native startups, not older SaaS teams adopting agents.

Large companies lead on adoption and buyer trust, while small teams compete on speed, focus and closeness to operations.

What If Your Competitor Is Another Small Team?

When your competitor is another small team, the advantage shifts from building to everything around the build. The HBR authors Wu and Anand foresee AI-first micro-teams and solo entrepreneurs matching a larger firm’s scope and speed with a fraction of its headcount. In Stripe’s survey of 200 SaaS platform leaders (March 2026), 57% were concerned about losing customers to AI-driven competitors, and 45% said AI investments had already lowered operating costs.

TSL’s reading: if building gets cheaper for everyone, the differentiators left are the ones a product idea cannot copy, such as customer relationships, trust and the quality of your oversight. None of the sources reviewed tests that directly.

How Should a Small SaaS Team Start With AI Agents?

Anthropic’s advice to builders is to start with the simplest solution and add complexity only when needed, and it says that can mean not building an agentic system at all. The Five-Gate Rollout turns that advice, and the failures in Sections 6 to 14, into a go-live checklist.

What Are the Five Gates?

The table shows, for each gate, the question to answer before go-live, the failure it prevents and the pass signal.

The Five-Gate Rollout: five checks an agent passes before it goes live, and keeps passing.

How Do You Measure ROI for a Small-Team Agent?

You measure ROI with your own before-and-after numbers, net of fees and oversight time. TSL’s checklist:

• Quality: audit a sample of the agent’s “resolved” cases yourself (Section 6).
• Escalations: count how often cases reach a person, and whether the handoff was right.
• Repeat contacts: track customers who return on the same issue.
• Time: measure hours with a timer, not recollection (Section 7).
• Cost: add fees and review time (Section 13).
• Errors: log incidents and what each cost (Section 14).

Example (illustration, arithmetic only): an agent drafts 400 replies a month. A person spends 1 minute on each review and 4 more minutes editing the 25% that need changes: 400 × 1 + 100 × 4 = 800 minutes, about 13 hours of oversight a month. That time belongs in the ROI calculation next to the vendor’s fee.

How Do You Spot Agent Washing Before You Buy?

Agent washing is vendors rebranding existing products as agents, and Gartner defines it as the rebranding of AI assistants, robotic process automation and chatbots without substantial agentic capabilities. Gartner estimates only about 130 of the thousands of agentic AI vendors are real, according to coverage of its release. TSL’s five questions for a vendor, using the Section 2 definition of an agent:

1. Does the product choose its own steps and tools, or follow a fixed path?
2. What can it access and change, and can you limit that?
3. When does it hand off to a person, and can you set the rule?
4. How does it define a “resolved” or successful outcome?
5. What does it report, and can you check those numbers yourself?

For the buy-versus-build call, see TSL’s 2026 build-versus-buy data.

What Should a Small SaaS Team Do Next?

The Five-Gate Rollout is the start: one job, least access, human approval, your own baseline and an escalation list.

AI agents help small SaaS teams on routine, well-defined work, and the evidence shows them changing how people work more than whether teams hire. Large companies keep the lead on adoption and buyer trust, so small teams compete on speed, focus and closeness to customers.

Start narrow, keep a person in the loop and measure your own results. To separate what is in production from what is still hype, read TSL’s guide to agentic AI in B2B SaaS.