Decision guides
Scope, value, buy-or-build, evaluation, approval and readiness.
Choose a starting point
Where should our business start with AI?
Start with one recurring workflow that has a named owner, a measurable current cost, outputs a person can check quickly, and information you are allowed to use. Rank a handful of candidates on value, feasibility, risk and readiness, then pick the one you can prove or disprove within a quarter — even if it isn't the biggest prize.
Read the guide →Establish value
How to calculate AI ROI honestly: hours released are not cash saved
Count hours released separately from cash removed: time only becomes money if it cuts overtime, avoids a hire or contractor, or turns into output someone will pay for. Subtract the ongoing costs — review time, licences, support and change management — then test the result across low, base and high cases and agree in advance what would make you stop.
Read the guide →Buy, configure or build
Should we buy, build, configure or do nothing about AI?
Work up a ladder of options and stop at the lowest rung that meets the need: do nothing, redesign the process, configure software you already own, add ordinary automation, buy a specialist product, build an AI-assisted workflow on a platform, and only then commission a custom agent. Each rung costs more to run and maintain, so the burden of proof rises with it.
Read the guide →Compare products
How to evaluate AI tools for your workflow, not the vendor's demo
Decide what the workflow needs before you look at products, then test each shortlisted tool on a set of real examples from your own work, scored against the same rubric. A good trial measures output quality, review time, data handling, admin controls, integration and cost at your real volume — not how impressive the demo felt.
Read the guide →Prepare procurement
Writing an AI implementation brief, and the questions to ask suppliers
Give suppliers a written brief covering scope, systems, information, constraints, acceptance criteria, open questions and your purchasing requirements, so every quote answers the same problem. Then ask each one for evaluation evidence, how they handle your data, who will actually do the work, how acceptance is tested, where support ends, how you can leave, and whether they earn fees from the products they recommend.
Read the guide →Handle company information
Can staff use AI with internal documents?
Sometimes, but approve a specific workflow's data flow rather than "AI" in general. Match the sensitivity of the information to the tier of tool and the contract terms behind it, exclude the categories that need more care, and record the decision so people know what is allowed.
Read the guide →Approve agent actions
What should require human approval for AI agents?
Decide per action, not per agent. Score each action on how reversible it is, how much damage a wrong one could do, and how confident you can be that it's right, then assign an approval pattern from draft-only to autonomous within limits. Back that with least-privilege access, tool allowlists, hard limits and an audit log.
Read the guide →Move beyond a pilot
Why AI pilots stall — and what production-ready actually requires
Most pilots stall because they were set up to demonstrate a tool rather than to pass a decision: no agreed acceptance threshold, no evaluation set, no owner for running it, and unresolved data, integration or security questions. Review the pilot against a production-readiness checklist, then make an explicit go, fix-then-go or stop decision.
Read the guide →Enable leadership
What should the board ask before investing in AI?
Ask for AI to be presented as a portfolio of specific workflow bets rather than a single programme, each with an owner, a measured baseline, the alternatives considered, honest costs, the main risks and the conditions under which it will be stopped. Then review the portfolio on a fixed cadence and move money towards the bets that are producing evidence.
Read the guide →Workflow guides
Specific business workflows, with what to automate, what to review and how to measure.
Activate existing tools
How to improve Copilot or ChatGPT adoption across a team
Stop promoting the tool and start agreeing how it is used in three specific workflows: the inputs, the prompt or template, the review step and where the output goes. Tie safe-use rules to your information classes, appoint a champion per team, and measure effort, quality and active use on those workflows rather than raw logins.
Read the guide →Reports and packs
How to automate recurring client reports and packs (and check AI-written summaries)
Automate the figures with ordinary tools — queries, templates and BI — and use a language model only to draft the narrative from those figures. Every number in the final pack must trace back to a source, and a named person checks the AI-written commentary for figure accuracy, unsupported claims and tone before it goes to a client.
Read the guide →Research and proposals
Faster research briefs and proposals from approved information
Yes, if the model drafts only from an approved, current content library and cites the source of every claim. Subject-matter experts review the draft, and the decision to bid, the pricing and the commitments stay with people.
Read the guide →Fix knowledge tools
Why does our internal knowledge assistant give bad answers?
"Bad answers" usually has several different causes — the right document wasn't found, the documents disagree or are out of date, the user can't see the source, or the question is outside what the assistant should answer. Classify real failures, build an evaluation set from real questions with expected answers and sources, and fix the cheapest causes first before considering a rebuild.
Read the guide →Maintain value after deployment
How to measure AI adoption — and when to retire an AI workflow
Give each AI workflow a named service owner and review its health on a fixed cadence: use on the intended workflow, quality against a re-run evaluation set, review effort, incidents, cost per run and maintenance burden. Keep, extend or retire it against criteria agreed in advance, not on enthusiasm or habit.
Read the guide →Looking for background?
The open GenAI and Agentic AI playbooks cover the concepts behind these decisions in 21 chapters and 10 languages.