45 Designs. One Pattern.

We've designed 45 AI agent architectures. Across accounting, law, recruitment, construction, property management, roofing, marketing, healthcare, software consulting, SaaS, logistics, HR, e-commerce, dental, and engineering.

The businesses ranged from 14 employees to 46. The revenue from £1.4M to £5.2M. The bottlenecks from engagement letter generation to knowledge capture to patient recall. The tools from Xero to Dentally to Manhattan Associates.

The technology varied. The industries varied. The people varied.

The pattern didn't.

Every working agent shares five characteristics. They showed up in the first design (Karen's engagement letters in Portland) and the most recent (Megan's patient recall in Cardiff). They showed up in SaaS and in roofing. In a law firm and in a dental practice. The consistency is the finding.

Every working agent targets a single, specific bottleneck. One workflow. One process. One named problem describable in one sentence.

Karen's engagement letters. James's revenue recognition. Sophie's churn visibility. Tom's client health. Claire's onboarding milestones. Graham's undocumented knowledge. Alison's support sentiment. Megan's patient recall.

Each one describable in a sentence. "Revenue recognition across 14 projects takes 1.5 days per month because three systems don't talk to each other." "720 patients are overdue for recall and the only outreach is a receptionist calling on Monday mornings." That level of specificity. That level of precision.

The agents that worked are the ones where the bottleneck was named before the technology was chosen. The businesses that came to us saying "we want AI" stalled. The businesses that came saying "this specific process is costing us £40,000 a year and we can describe exactly why" built something in two weeks.

The 3-Touch Test exists because of this pattern. Touch 1: "Which single workflow loses you money or time every week?" Singular. Specific. If you can name it in one sentence, you can build for it. If you can't, the scope needs narrowing before anything else happens.

In 44 of 45 designs, the data the agent needed was already sitting in the business's existing systems. Zendesk held the ticket sentiment. Dentally held the recall dates. Teamwork held the project milestones. Mixpanel held the usage data. The WMS held the delivery performance. SharePoint held the contracts. BreatheHR held the onboarding records.

The agent's job was never to create data. It was to connect, read, and act on data that already existed in systems that didn't talk to each other.

The human was the bridge. Karen between Xero and the template library. James between Teamwork, SharePoint, and Xero. Sophie between Mixpanel, Intercom, Stripe, HubSpot, and Typeform. Tom between the WMS, TMS, Sage, and Salesforce. Helen between Dentally and a printed spreadsheet. Graham between his 34 years of experience and whoever was standing in his doorway.

The agent replaced the bridge. Not the data. Not the systems. Not the people. The bridge. The invisible, unnamed, untracked work of moving information between systems that should have been connected. The invisible 30% that consumed the most capable person's time in every business we examined.

The one exception (the 45th design, Graham's knowledge capture) was the only case where the data didn't exist in a system. It existed in a person's head. The agent's first job was to get it out of his head and into a system. Every subsequent query hit the system. The principle still holds: the agent connects data to the people who need it. It just had to create the data store first.

If your bottleneck involves someone copying data between systems, cross-referencing multiple sources, or synthesising information from tools that don't connect, the data for your first agent is almost certainly already in your stack. You don't need new data. You need the data you have to be connected.

Every working agent handles the reading and leaves the caring to a human. Every one.

Reading: classification, scoring, aggregation, matching, retrieval, monitoring, calculation, pattern detection. The work that requires processing information at volume, consistently, across systems. AI does this extraordinarily well. Haiku classifies 3,200 support tickets for $1 per month. The maths is not close.

Caring: judgment, empathy, decisions, relationships, exceptions, context. The work that requires understanding why the information matters, what to do about it, and how the outcome affects a person. AI does this badly. A sentiment score of negative 0.8 tells you the customer is angry. It doesn't tell you the customer's child's birthday present arrived damaged and the issue is the birthday, not the product.

The split across every Blueprint: roughly 80% reading, 20% caring. Alison's agent reads 3,200 tickets. Alison decides what to do about the delivery partner. Rachel's agent applies 47 expense rules. Rachel handles the genuine grey zones. Graham's agent retrieves documented answers. The engineers apply professional judgment. Helen's agent sends SMS sequences. Helen calls the patients who need a human voice.

Failed implementations tried to push past this line. The chatbot that handled customer complaints without human involvement. The AI writer that replaced the content team. The automated decision-maker that approved expenses without review. Each one asked AI to care when it can only read. Each one failed.

The line is the design principle. Know where it is. Design the agent on one side. Design the human role on the other.

Every working agent has a named person who reviews, approves, or acts on the agent's output.

Karen reviews the engagement letters. James reviews the recognition schedule. Rachel reviews the flagged expenses. Claire reviews the onboarding dashboard. Megan reviews the recall report. Helen calls the patients who didn't respond. Tom reviews the flagged accounts. Alison reviews the escalation-risk tickets. Graham reviews the knowledge articles.

The human's role changed. Karen went from producing 412 letters to reviewing them. James went from 1.5 days of cross-referencing to 2 hours of reviewing a pre-calculated schedule. Helen went from calling 47 patients to calling 15. The work shifted from doing to reviewing. The human was promoted from "the system" to "the expert who makes the system better."

This is the design choice that builds trust. The agent proposes. The human disposes. For the first 30-60 days, the human reviews everything. After trust builds, they review exceptions only. The transition is gradual, driven by evidence, and controlled by the human. Nobody is forced to trust the system before the system has earned trust.

Agents deployed without a named human owner get turned off within 60 days. Every time. Because the first error with no human catch destroys confidence permanently.

Running costs across the 45 designs: £24-£280 per month. The median is approximately £180.

The problems they address: £9,000-£640,000 per year in direct costs, lost revenue, or risk exposure.

The ratios are consistently absurd. A £38-£65 per month agent recovering £38,000-£52,000 in recall revenue. A £50-£130 per month agent capturing 94 years of institutional knowledge before it retires. A £21-£47 per month agent reading 3,200 tickets that a £165,000-per-year team was synthesising from memory.

The Claude API costs are particularly striking. Classifying 3,200 support tickets: $1 per month. Running a knowledge retrieval system for an engineering consultancy: $4 per month. Generating weekly sentiment reports: $0.10 per month. The unit economics of AI classification are so favourable that the cost of the agent is dominated by the hosting, not the intelligence.

The build cost is the real investment: £1,800-£7,000 for most agents when hiring a developer. Or free, if you build it yourself using the guides we publish with every Blueprint.

This is the number the AI industry prefers you don't see. Because an SMB owner who realises their first agent costs £180 per month is an SMB owner who finds the £35,000 discovery engagement rather difficult to justify. Every Blueprint we publish includes the full cost breakdown. Build cost, monthly running costs by component, Claude API token-level detail. Total transparency. Because the maths is the argument, and the maths doesn't need a vendor deck to make it.

One bottleneck, precisely named.

The data already exists.

80% reading, 20% caring.

A named human in the loop.

Cost that's embarrassingly low.

Five characteristics. Present in every working agent. Absent in every failure.

This is what 45 designs taught us. Not that AI is magic. Not that automation solves everything. Not that agents replace people. That if you name one specific bottleneck, connect the data you already have, let the agent read and the human care, keep someone in charge, and spend £180 per month instead of £50,000, you get a system that works. Quietly. Reliably. For years.

Karen still reviews the engagement letters. James still reviews the recognition schedule. Sophie still makes the retention calls. Tom still has the client relationships. Claire still judges how new hires are settling in. Graham still applies 34 years of professional judgment. Helen still calls the patients who need a human voice. Megan still runs the practice.

Nobody was replaced. Everyone was freed from the work that shouldn't have been theirs in the first place.

The full archive is free at adai.news. Forty-five designs across fifteen industries. Your industry is probably already in there. The architecture, the build guide, the cost breakdown, the failure modes. Everything you need to build the first one yourself.

Subscribe to AdAI News for the next one. And the one after that. The series continues.

by SP, CEO - Connect on LinkedIn
for the AdAI Ed. Team