AI process automation — built, tested, and handed over running
One workflow, connected to the systems you already use, tested against your own historical work, with a person reviewing the output before it is used.
Four things that are true of every build
We start from your real examples.
Tested against work your business has actually done, not sample data.
A person always checks.
Review steps are designed into the sequence, not added at the end.
It connects to what you already run.
Nothing is replaced.
Nothing depends on us afterward.
Documentation, training, and a support window — and nothing running on a credential only we hold.
Three sizes of AI workflow implementation
| Lite | Standard | Advanced | |
|---|---|---|---|
| Scope | One recurring task | A full multi-step sequence | Complex or commercially significant, or two linked sequences |
| AI-assisted steps | 1 | 2–3 | 3–5 |
| Connected tools | 1–2 | 2–4 | 3–6 |
| Testing | 10–20 real cases | 30–50 real cases, batch scored | 50+ cases plus deliberate testing of unusual ones |
| Rollout | Handover session | Team training | Phased rollout with adoption support |
| Support after go-live | 7 days | 14 days | 30 days |
| Timeline | 2–3 weeks | 4–6 weeks | 6–10 weeks |
Lite — one recurring task handled by AI
One task that recurs weekly or more often, where a single step requires reading, writing, sorting, or summarizing, touching one or two systems.
The input arrives as free text, speech, or documents that vary every time. A rule cannot read a supplier's email and decide what it is about.
Claude or OpenAI for the reading and drafting step · Make or n8n to connect · your existing storage, project, or CRM system.
- 01Recording lands in shared folder
- 02AI produces structured notes: decisions, actions, owners, dates
- 03Draft appears in the project system, marked for review
- 04A person checks and approves
- 05Sent
Meeting recordings into client-ready notes
An architecture practice records client and site meetings. Someone spends around six hours a week listening back and writing up decisions, actions, and who owns them.
Distinguishing a decision from a discussion, an action from an aside, and attributing each correctly — judgment about meaning that no rule can encode.
The recording lands in shared storage. Claude produces a structured note with decisions, actions, owners, and dates in the practice's own format, flagging anything ambiguous rather than guessing. The draft appears in the project system for the architect to approve.
Scope: one AI step, two connected systems, tested against 15 historical recordings. Two to three weeks.
Where the whole sequence around this needs rebuilding rather than one step, that is Standard.
Standard — a full sequence rebuilt with AI
A sequence with several steps involving more than one person, spanning two to four systems, where two or three steps involve reading, writing, sorting, or checking.
More than one point in the sequence requires interpreting something written by a human — an inquiry, a document, a conversation — and producing something structured from it.
Claude or OpenAI for extraction and drafting · Make or n8n for the sequence · HubSpot or your CRM · document storage · email.
- 01Inquiry arrives — email, form, phone note
- 02AI extracts requirements, scope signals, and what is missing
- 03Structured record created in the CRM
- 04AI drafts a scoped proposal from your template library
- 05A person reviews, adjusts, and sends
- 06Follow-up scheduled; response classified when it arrives
Inquiry to scoped proposal
A commercial fit-out contractor receives inquiries as long, unstructured emails, often with drawings attached. Each one is read by an estimator, summarized, entered into the CRM, and turned into a proposal — a sequence of two to three hours before anyone has priced anything.
Reading a discursive email and identifying scope, timescale, site constraints, and what has not been specified; then producing a first-pass proposal in the firm’s own structure and language.
The inquiry is read and reduced to a structured record — scope, location, indicative value, timescale, and an explicit list of what is still unknown. That record populates the CRM. A proposal is drafted from the firm’s template library using their own wording. The estimator receives a draft with the open questions listed, rather than a blank page.
Scope: three AI steps, four connected systems, tested against 40 historical inquiries. Four to six weeks.
Supplier invoices into a coded ledger
An accountancy practice receives client bookkeeping in whatever form it arrives — photographs, PDFs, forwarded emails, occasionally a shoebox. Junior staff extract the details and code each item.
Reading documents with no consistent layout, and deciding the correct account code from a supplier name and description that vary between clients.
Documents arriving in a monitored inbox are read and the supplier, date, net, tax, and total extracted. A code is proposed based on the practice’s own historical coding for that client, with a confidence signal. Anything below the threshold routes to a person with the reason. Approved entries post to Xero or QuickBooks.
Scope: two AI steps, three connected systems, tested against 50 historical documents including deliberately poor scans. Four to six weeks.
Where the sequence is genuinely one task, Lite covers it for less. Where it spans five or more systems or several teams, that is Advanced.
Advanced — complex and business-critical AI workflows
A sequence where several steps involve judgment, drafting, checking, or analysis, spanning three to six systems, with multiple roles involved and output that reaches clients or drives decisions.
The sequence carries genuine variation, and the cases outside the normal pattern have to be identified and routed rather than forced through.
Claude or OpenAI for classification, extraction, and drafting · n8n or Make for the sequence and routing · your project, finance, document, and CRM systems · a reporting layer where needed.
- 01Documents and data arrive from several systems
- 02AI classifies, extracts, and cross-checks against your standards
- 03Anything outside the pattern routed to a named person, with the reason
- 04AI drafts the output: report, response, assessment, or summary
- 05Role-based review — the right person checks the right part
- 06Published to the client-facing system, with a record of what was decided and why
Candidate screening and shortlist notes
A recruitment agency receives large volumes of applications per role. Consultants read each CV against the brief, form a view, and write shortlist notes for the client.
Reading a CV written in any format and assessing it against a specific brief; recognizing that unrelated job titles describe relevant experience; and writing a shortlist note in the consultant’s own voice.
Applications are read against the role brief and scored on the criteria the agency actually uses, with the evidence for each score quoted from the CV rather than asserted. Anything borderline is flagged for a consultant rather than rejected. Shortlist notes are drafted in the agency’s house style. Every rejection remains a human decision — the workflow ranks and evidences, it does not decline anyone.
Scope: four AI steps, five connected systems, deliberate testing against historical applications including ones the agency judged marginal. Six to ten weeks with a phased rollout by desk.
Policy renewal review
An insurance brokerage reviews client policies ahead of renewal. Each review means reading the existing policy, the client’s updated details, and the market alternatives, then producing a recommendation.
Comparing policy wordings that express the same cover differently; identifying where a client’s circumstances have changed in a way that affects cover; and drafting a recommendation with reasoning a client can follow.
The current policy, claims history, and updated client information are read and reduced to a structured comparison. Changes in exposure are identified and flagged. A renewal recommendation is drafted with the reasoning set out. Any recommendation involving a coverage change is reviewed by a broker before it reaches the client, and anything the workflow cannot assess confidently is routed with the reason rather than assumed.
Scope: four AI steps, six connected systems, deep testing against historical renewals including disputed ones. Six to ten weeks.
Delivery and margin monitoring
A marketing agency runs client work across a project system, a time tracker, and accounting software. Whether a job is running below margin becomes clear when it is invoiced, which is too late to act.
Reading project updates, notes, and client correspondence to identify scope movement that has not been formally recorded — the phrase in an email that means the job has grown but no change order was raised.
Activity across the three systems is assembled weekly. Time booked is compared against fee. Written updates and correspondence are read for signals of scope movement. The output is a weekly view of which jobs are drifting, which are running below expected margin, and which have grown without a change order — with the evidence quoted rather than asserted.
Scope: three AI steps, five connected systems, tested against twelve months of historical project data. Six to ten weeks.
Where a sequence spans two to four systems and one team, Standard covers it in four to six weeks.
Outside this engagement
- Replacing or rebuilding core business systems.
- Custom software built from scratch.
- Connections to systems with no viable way to connect.
- Output used without human review.
- Unlimited scope changes mid-build.
Where to start
If the task is already identified and agreed internally, we can scope an implementation directly. If it has not been settled, an AI readiness audit identifies and ranks the candidates — and usually finds more than one. Where several teams are involved, the Deep Audit includes a build-ready specification for the top recommendation.