AI for professional firms and agencies

The billable hour is not where the time goes.

Across the agency engagements in SUPALABS’s data, the average fee-earner loses 12–18 hours a week to non-billable admin: timesheets, status updates, scope-change paperwork, onboarding documents, the first draft nobody wants to start (SUPALABS engagement data, 2024–2026). In a commercialista’s office it is the scadenzario and the antiriciclaggio file; in a law firm the conflict check and the precedent search; in an engineering studio the tender documents. None of it is the judgement the client pays for. All of it is a queue.

  • 30–50% faster deliverable turnaround in research-heavy work, concentrated in research and first draft (SUPALABS engagement data)
  • Built on top of the practice management and document tools you already run
  • The professional signs everything that leaves the firm
Six queues, six roles

Put an agent on the queue, not on the org chart.

Each role owns one recurring queue. It drafts, it shows its evidence, and a named person signs before anything consequential moves. The last column is the sector’s AI Boundary Map: whether the step is mostly a rule, mostly a model, or stays with a person.

RoleWhat it draftsWho signsWhat it readsMostly
Client intake and onboardingThe onboarding file from the documents the client sent: identity and beneficial-owner checks assembled, the engagement letter drafted from the firm’s template, missing items requested.The partner, before the engagement letter goes out.Email attachments, the practice management system, the firm’s templates.a model
Research and first draftA first draft from the firm’s own precedents and the sources you approve, every claim linked to where it came from, and the parts it could not source marked as such.The professional, who edits rather than starts.The document store, the precedent library, approved databases.a model
Time and scope-change captureThe timesheet reconstructed from calendar, email and ticket activity, and a scope-change note drafted whenever work appears that the engagement letter did not cover.The fee-earner, who confirms rather than remembers.Calendar, email, the project tool.a model
Deadline calendarEvery deadline per client from the mandates and the regulatory calendar, with the documents each one needs and who has not yet sent them.The office, weekly.The practice management system, the regulatory calendar.code and rules
Client status reportingThe weekly status per client from the project tool and the inbox, in the firm’s voice, with open questions listed rather than glossed.The account lead.The project tool, email, the CRM.a model
Billing and work in progressThe invoice from approved time and scope, and the WIP position per client with anything unbilled for more than thirty days flagged.The partner.Approved timesheets, the engagement letter, the ledger.code and rules
The determinism ratio

How much of this is actually AI.

This is the one sector where the model does most of the drafting, because the raw material is text: documents, precedents, calendars, inboxes. Four of six queues need one. That makes the evaluation suite the artifact that matters most here, and it is why the professional signs everything that leaves the firm. In research-heavy consulting engagements SUPALABS has measured a 30–50% acceleration in deliverable turnaround, concentrated almost entirely in research and first draft, and almost none of it in the judgement a senior professional still makes.

2 / 6 queues mostly rules · 4 / 6 with a model

Proof

No client names. Real patterns.

Every engagement produces the same five documents, whatever the sector. What we can show is anonymised: the pattern, the number, and where it applies.

12–18 hours per week, per fee-earner, lost to non-billable admin across agency engagements; 30–50% faster deliverable turnaround in research-heavy consulting work (SUPALABS engagement data, 2024–2026).

Related case studies

An AI Content Engine That Cannot Publish Without a HumanEvery AI writing tool promises you stay in control. We built one where that is a missing function rather than a marketing line — the publish tool does not exist in the automation surface at all. Read the case study →
A Venture Engine Where the Model Cannot Talk Its Way Past the GateAsk a language model to judge a business idea and it will argue itself into a yes. We built a screening engine where the verdict is derived by code, and the cross-examination runs on a different model. Read the case study →
Eight Specialist Agents That Screen a Data-Centre Site in One RunPre-due-diligence on a candidate site used to mean weeks of separate specialist reviews in incompatible formats. We built a swarm of eight domain agents that queries real public data and assembles one scored, sourced report. Read the case study →
All engagement data, with sources →
How authority is earned

Three tiers, set per task type, and you set the thresholds.

Every automated task runs at one of three levels of authority. The tier is set per task type rather than per system, so the same workflow can draft one step, wait for approval on another, and act alone on a third. A task moves up only when the evaluation suite shows the pass rate you defined as the threshold, and it moves back down the moment the monthly report shows a regression.

1
DraftedThe system prepares the work and a person finishes it. Every draft carries its evidence: the source documents, the rule or model that produced it, and its confidence. This is where every task starts, and where shadow mode runs before anything is shown to the team at all.Gated by the Exception Ledger — the task is not automated at all until its real deviations are written down.
2
ApprovedThe system prepares the complete action and a named person authorises it before anything consequential moves. The approval, the approver and the inputs go into the decision log.Gated by the evaluation suite — promotion from Drafted needs the per-step pass rate you set, measured against the golden dataset.
3
AutonomousThe system acts and the team audits afterwards, on a sample or on the exceptions. Any irreversible step stays at Approved regardless of accuracy, because a reversible mistake is a cost and an irreversible one is a liability.Gated by the monthly accuracy report — a regression alert sends the task back down a tier until the pass rate recovers.

Nothing is promoted on our say-so. The thresholds are yours, the report that tests them is a named line item, and the decision log shows every action that was taken at every tier.

The full method: five artifacts, three tiers, four rungs →
How an engagement runs

Four rungs. You can stop after any of them.

0
Qualification call — free, 30 minutesA straight answer on whether there is anything here worth mapping. No deck, no proposal, no follow-up sequence.
1
Mapping Sprint — five days, with the people who do the workPaid discovery. Produces the Exception Ledger, the AI Boundary Map, the evaluation plan and either a fixed-price build quote or a written no. If it shows you nothing you did not already know, you do not pay.
2
Build — six weeks to productionScoped from what the sprint found. Majority deterministic software on top of your existing systems, every task starting at Drafted, shadow mode before anything acts.
3
Run — ongoing, and optionalMonthly accuracy report against the golden dataset, regression alerts, cost-per-run tracking. Stop the service and keep the system.
From the blog

The problems, one at a time.

AI Automation for Agencies: The 2026 Operator's PlaybookAI automation for agencies, decoded: the 8 workflows to automate first, the 2026 tool stack, and what the numbers actually look like at a 25-person shop.
AI Tools for Consultants: The 7 We'd Actually Pay For (2026)The 7 AI tools for consultants we'd actually pay for in 2026 — real pricing, client confidentiality traps, and the workflows that hold up under partner review.
AI Workflow Automation Agency: How to Hire One in 2026How to hire an AI workflow automation agency in 2026: 6-dimension framework, pricing models, red flags, paid pilot test, and questions to ask.
AI for Italian Marketing Agencies - Automate Reports, Content and Clients in 2026In-depth guide: AI for Italian Marketing Agencies - Automate Reports, Content and Clients in 202...
AI for Legal Firms: Document Automation and Case Management RevolutionComprehensive guide to AI implementation in legal practices. Document automation, case management, legal research tools, and ROI analysis for law firms seeking competitive advantage.
Legal Document Automation 2026: Cut Legal Costs by 70% with AI Contract ToolsComplete 2026 guide to automating legal documents for SMEs. Compare PandaDoc, Juro, Gavel, and Ironclad. NDA generators, contract automation, e-signatures, and compliance tracking without expensive lawyers.
AI Lead Generation for Accountants and Tax Consultants in Italy in 2026AI lead generation for Italian accountants and tax consultants: new business registration signals, compliance deadline targeting, LinkedIn authority for commercialisti, CPL benchmarks for accounting services.
AI Lead Generation for Architects and Engineering Firms in Italy in 2026AI lead generation for Italian architects and engineering firms: building permit signals, public tender monitoring, BIM project tracking, referral automation. CPL benchmarks for technical services.
Workflow Automation Specialist: 2026 Hiring GuideWhat a workflow automation specialist actually does, current rates, the skills that matter in 2026, and how to tell a real expert from a Zapier hobbyist.
FAQ

Frequently asked questions

Will a model write advice under our name?No. The role produces a first draft with every claim linked to its source, and the professional edits and signs. Drafts are the Drafted tier by definition, and client-facing advice never moves above Approved, whatever the accuracy report says, because the signature is the product.
What about client confidentiality and professional secrecy?Self-hosted tooling on EU infrastructure for sensitive matters, data excluded from model training, full data-processing agreements, and a decision log that shows what was read and by which step. Where a matter cannot be processed at all, it stays with a person, and the Exception Ledger says so.
We are a firm of eight. Is this for us?If one of the queues on this page costs you a person-day a week and you can name who owns it, yes. If the problem is arithmetic, a well-built spreadsheet wins and we will say so on the call. If you are large enough to run your own delivery team, hire it.
How do you measure whether the draft is good?With a golden dataset built from your own past work: the drafts your professionals would have accepted, scored step by step. That is the evaluation suite, quoted as its own line item. The monthly accuracy report runs against it, so you know if the drafts are getting worse before your clients do.
Do you replace our practice management software?No. We build on top of it and of the document tools you already use. The queues on this page read from those systems and write drafts back into them. That commitment is in writing from the start.
What does it cost?The thirty-minute qualification call is free. Everything after it is quoted per engagement, because the scope is a finding rather than an assumption. If the Mapping Sprint shows you nothing you did not already know, you do not pay for it.
Other sectors

The same method, other queues.

Wholesale & B2B distributionEvery order is a promise. Most of the work is checking whether you can keep it.
ManufacturingThe order arrives forty different ways. The line needs it one way.
All sectors →

Thirty minutes to find out whether there is anything worth mapping.

The call is free, and we will tell you if the answer is no. Bring one queue that costs someone an evening a week.