13. How It Changes by Sector

Every CEO I meet asks the same question in the first ten minutes: "Fine, but in my sector?" The honest answer has two halves. The method does not change: the audit, the Conditions of Satisfaction, the first low-risk agent with a fast return, the steps. What changes is where the time is lost today, which agent to place first, which number to watch, and which risk — legal or reputational — must be designed in before anything goes live.

This chapter gives you both halves. First the part that never changes, in a page. Then a grid with one row per sector, so you can find yours and read what to do first. Then three closer looks — retail and e-commerce, manufacturing, professional services — where I have the most direct experience. By the end you will be able to name the first agent for your company and the number it must move.

What never changes

Four things hold in every sector, and skipping any of them is the most common cause of failure I see.

The audit comes first. Before any agent, the photograph: which processes absorb the most hours, what they cost, where the data lives, where errors and delays are born. Without a baseline there is no demonstrable result, and without a demonstrable result there is no second step.

The first agent is low-risk and fast. Not the most ambitious use case; the one that touches no critical decision, has clean data, and shows a measurable result in four to eight weeks. The purpose of the first agent is not the return: it is to create the proof, the first supervisor, and the appetite for the second.

Every agent moves one of three numbers. Lower costs, higher revenue, or more capacity without hiring. An agent that cannot be tied to one of the three with a CoS is an experiment, and experiments are for laboratories.

The steps are the same. Audit, first agent, first AI Team, mixed organization — chapter 5. Sectors differ in the pace, not in the sequence.

The grid

Read your row, then read the neighboring rows: the agents in a sector next to yours are usually your second and third.

Data point. In a compilation of anonymized 2026 cases, the first agent produced comparable results across very different sectors: a lender cut loan-document processing from 4.2 to 1.1 hours per file (−38% labor cost); a B2B software company cut sales admin from 65% to 28% of reps' time (+24% qualified pipeline); a regional healthcare network cut prior authorization from 5.3 to 2.1 days (−52% backlog); a $400 million manufacturer cut supply-chain exception resolution from 3.8 to 2.1 days (−44% staff time). Different sectors, same pattern: a routine process with clean data and a human on the exceptions. Source: Deployed Labs, AI Agents Business Results & ROI Case Studies for 2026.

Retail and e-commerce

This is the sector where the market has already changed on its own, and where the first agents pay back fastest. Customers arrive through AI assistants (chapter 6), expect answers in minutes, and abandon at the smallest friction. Recommendations already drive a large share of online revenue — up to 35% of sales at Amazon — and AI personalization lifts revenue by 10-15% on average.

Where to start. Two agents, in this order: recovery of open carts and quotes, and routine support (order status, returns, sizes, availability). Both are low-risk, both are measured in four weeks, both free the people who today answer the same question fifty times a day. The third is product-information governance: keeping catalog data complete and consistent for the assistants that now decide whether you are proposed.

The numbers. Conversion rate, cost per order, share of revenue from existing customers. In inventory-heavy retail, add capital tied up in stock: AI inventory optimization cuts overstock write-offs by 14% and stock-outs by 11%.

The risk. Klarna's story is the lesson: its assistant handled 2.3 million conversations in its first month, then the company rehired people because cost-driven automation had lowered quality. Two-thirds of inquiries still go to AI; the "moments that matter" go to people. Design that split from day one, and never hide the agent from the customer.

The organization. The e-commerce manager becomes the head of a commercial AI Team; customer-service operators become supervisors of exceptions; the external agency moves from execution to governance, or leaves.

Manufacturing

Industrial SMEs have the best-documented returns and the lowest adoption: predictive maintenance cuts unplanned downtime by 30-50% and extends equipment life by 20-40%, with payback in 6-18 months — yet only 27% of manufacturers use it actively. The gap is the opportunity.

Where to start. Maintenance, if you have machinery with sensors or can add them; technical quoting, if your bottleneck is the engineer who prepares offers; supplier follow-up, if delays come from the supply side. All three are routine, data-rich and far from the product's safety-critical decisions.

The numbers. Unplanned downtime, quote turnaround, on-time delivery, capital in stock. In the export-heavy SME from chapter 7, the number was response time to foreign requests: from a week to a day.

The risk. Acting on bad data. An agent that schedules maintenance on a faulty sensor, or reorders on a wrong forecast, costs real money. The playbook thresholds from chapter 7 exist for this: inside the thresholds the agent acts, outside it proposes.

The rules. From August 2028 AI embedded in machinery and regulated products falls under the EU's high-risk regime. Agents that plan, quote and monitor do not; agents inside the product do. Keep the two apart in the inventory of chapter 12.

Professional and B2B services

Law, accounting, consulting, engineering, marketing services: the sector where AI changes the business model itself, because it sells hours and agents compress hours. Professionals expect to free around 240 hours a year each; document review falls by 60-70%; firms with a visible AI strategy are 3.9 times more likely to see a return; and 71% of legal clients already prefer flat fees.

Where to start. Document review and first drafts under professional supervision, then administration and billing. The first agent must never produce advice that reaches a client unreviewed: the professional signs, the agent prepares.

The numbers. Hours per matter, matters per professional, share of revenue on fixed fees, realization rate. The strategic number is the second: capacity per professional up 10-40% without hiring is the whole economic case.

The risk. Errors in advice and breaches of confidentiality. Both are governed by contract and by design: no client data in tools without a data-processing agreement (chapter 12), and a review gate before anything leaves the firm.

The business model. This is the sector where chapter 4 bites hardest. When a task takes a fifth of the time, hourly billing gives the saving to the client; fixed fees keep it in the firm. Nearly a third of in-house legal teams say they will reconsider outside firms that cannot show AI-enabled value within a year. The move to productized, fixed-price services is not optional; the question is who moves first.

Example. A twelve-person accounting firm placed its first agent on the monthly closing of client books: collecting documents, matching transactions, preparing the draft the accountant reviews. Time per client fell from six hours to under two; the firm took on forty new clients without hiring and moved the service to a fixed monthly fee. The two junior accountants who did the matching became supervisors of the agent and advisors to clients. (Case anonymized.)

The other sectors, in brief

Healthcare has the clearest evidence and the strictest rules. Ambient documentation saved 15,791 physician hours across 2.5 million encounters in one health system, with 84% of physicians reporting better patient communication; independent studies find more modest but consistent savings. Start with documentation and administrative workflows, never with clinical decisions; everything touching patient data or diagnosis is high-risk in every jurisdiction.

Hospitality and travel is a revenue-management story: independent hotels using AI pricing report RevPAR gains of 8-20%, and 78% of chains already use AI while only 41% of independents do. Start with pricing and guest messaging; watch RevPAR and direct-booking share; keep the human at the moments guests remember.

Logistics, construction and financial services share a pattern: the first agent handles documents and exceptions, the number is days-to-resolve, and the constraint is regulatory — customs and transport paperwork, contractual liability on technical documents, or the credit and model-risk rules that put finance at the top tier of every framework.

What to take away

The method is the same everywhere; the first agent, the number and the risk are sector-specific. Retail starts with recovery and support and watches conversion; manufacturing starts with maintenance and quoting and watches downtime; professional services start with document review and watch capacity per professional — and must change how they bill. In every row of the grid the first agent is routine, data-rich and far from the decisions that carry liability.

Your to-do list.

1. Find your row in the grid and write down the first agent and the number it must move.
2. Read the two neighboring rows: they are your second and third agents.
3. Check the "rule or trust constraint" column against the inventory of chapter 12 before the first agent is designed.