Since the start of this year I have rebuilt how ThinkTribal designs, assures and takes products to market around agentic AI, working hands-on at the command line rather than directing anyone else to do it. One rule governs all of it. Every automated output meets the same evidence standard as consultancy work a board signs off. A machine that writes for my company shows its sources, and where no source exists, it writes nothing.
Most organisations bought AI as an assistant and got plausible text. I wanted production infrastructure, which meant governing it like production infrastructure: defined inputs, quality gates, and an audit trail on every word.
The method survives its author
Consultancies have always sold methods, and methods have always lived in people. When the person leaves, the method goes with them, whatever the wiki says. This year I codified twenty-five of ours into reusable skills, each one a written method with defined inputs, evidence rules, quality gates and branded outputs, packaged as an installable bundle a colleague can run without me in the room.
A skill in this form is closer to an instrument than a document. Nobody re-reads a methodology under deadline pressure, and everybody runs the tool in front of them. The method executes the same way on the hundredth run as on the first, and it will keep executing after I have moved on to something else. Institutional knowledge normally walks out of the building on somebody's last day. This version stays.
The standard is already a specification
The same discipline took on regulated assurance. Data Protection Impact Assessments built to the ICO's seven-step method and Article 35(7) of UK GDPR. Business continuity and disaster recovery reviews aligned to ISO 31000, with recovery time and recovery point analysis behind them. White-hat security and access-control reviews. Benefits realisation producing dependency networks, benefit profiles and realisation plans rather than a spreadsheet of hopeful numbers.
None of this required inventing a method. Regulators and standards bodies published these methods years ago, and they read today exactly like specifications waiting for an implementation. What organisations lack is the capacity to execute them consistently, because a human working a seven-step method under deadline pressure skips step four. A codified method cannot. Automation does not water the standard down. It is the first thing I have seen that makes the standard enforceable.
The engineering layer got the same treatment: end-to-end and regression test audits, interactive UAT checklists carrying a keep, demote or cut verdict alongside pass and fail, site-wide consistency reviews, and technical debt worked down under branch-protected workflows. Quality stopped being a phase and became a set of running instruments.
Nothing is invented to fill a gap
The part most people assume cannot be automated safely is the part facing outward. ThinkTribal's go-to-market now runs on autonomous engines: a LinkedIn audience engine and a daily comment radar, working from a persistent memory of performance, watchlist, ideal customer profile and playbook, with measurement weighted towards the customers we actually want and defamation safeguards on any named organisation.
The sales automation for ScoreView and TenantSafe produces board exposure briefs, executive decks, one-pagers and outreach emails in which every claim traces back to a dated, sourced record. No invented return-on-investment figure. No capability the product does not have. No contact name the record does not hold. Where the evidence runs out, the document says less.
A machine-readable brand and tone system holds every artefact, from Word reports and slide decks to dashboards and video, to the same visual and verbal standards without manual policing. Even the product videos are generated from structured showcase evidence, with the provenance of every frame recorded.
Generation was never the hard part. Any model produces fluent prose about anything. Provenance is the engineering, and the reason most organisations dare not let AI near a customer is that they built the generation, skipped the evidence discipline, and discovered the gap the expensive way, in public.
What a board should ask
The commercial shift underneath all this was moving ThinkTribal from selling software to selling independent, board-ready assurance ahead of external scrutiny. The playbook holding that position is machine-readable, and it is the canonical input every automated engine reads before it writes a word. Strategy in this shape stops being a deck nobody opens and becomes a constraint nothing can ignore.
If you sit on a board being shown an AI programme, ask one question of it. Ask what any given output can prove: where its claims came from, which record they trace to, and what happens when the evidence is not there. A programme built as an assistant will answer with fluency. A programme built as infrastructure will answer with sources. The difference is worth more than every demo you will see this year.