Direct answers to
practical questions.

Begin with a business task or decision worth improving. Describe the current baseline, the people affected and what a useful change would look like. Then examine readiness: knowledge, systems, ownership and the ability to adopt a new way of working.

Choose a bounded starting point with an owner and a way to evaluate the result. The transformation approach in the AI Lab provides a sequence for structuring that discussion.

SOURCES AND FURTHER READING: AI Lab: transformation approach ↗ · From CTO to CTAIO ↗

Treat the deployment model as a requirements decision. Clarify which information the system will use, the access boundaries it must respect and the operational capabilities available to support it.

Compare the options against those requirements, including quality, latency, total cost and responsibility for running the system. Private deployment alone does not answer every question about knowledge quality, governance or usefulness.

SOURCES AND FURTHER READING: The Enterprise AI Stack Explained ↗ · NIST AI Risk Management Framework ↗

A useful strategy connects a business priority to a set of choices: what to pursue, what to defer, who owns the work and how progress will be assessed. It should make the next decision easier.

Keep technology choices connected to the workflow and the organisation adopting them. The point is a capability the business can use and improve, with evidence to guide further investment.

SOURCES AND FURTHER READING: AI Lab: transformation approach ↗ · Explore the expertise ↗ · Service: AI Strategy Advisory ↗

Agents introduce the possibility of software taking a sequence of actions toward a goal, rather than waiting for a person to direct each step. The design questions therefore extend to permissions, completion evidence and recovery.

Anthropic’s distinction between predictable workflows and flexible agents is a useful starting point. A product team can ask which user tasks justify that flexibility, and which are better served by an explicit process.

SOURCES AND FURTHER READING: Agentic AI and software development ↗ · Building effective agents ↗

Measure against a baseline you recorded before the system arrived, on the workflow it serves: cycle time, cost per task, quality and rework. If AI is not tied to margin, cost, risk or revenue movement, it is exploration — worthwhile, but not yet a return.

The discipline is to write a value case specific enough to fail, then evaluate honestly, including the failures. Most credibility problems with AI programmes trace back to success criteria that were never defined.

SOURCES AND FURTHER READING: From Pilots to Capability ↗ · The AI Readiness Assessment ↗

For most enterprise knowledge needs, retrieval comes first. RAG keeps answers current, auditable and permission-aware, because the system cites documents it retrieved under the asker’s identity. Fine-tuning changes behaviour and style; it is rarely the right tool for facts that change weekly.

The two combine well: retrieval for knowledge, light tuning for tone or format. But if you can only do one thing this quarter, improve the condition of the knowledge you retrieve from — ownership, currency, deduplication and permissions.

SOURCES AND FURTHER READING: Enterprise Knowledge Is the Real AI Advantage ↗ · Retrieval-Augmented Generation (paper) ↗

Usually the honest answer is: not in the way the phrase suggests. “Private” is a spectrum — contractual API terms, dedicated cloud tenancy, VPC deployment, full self-hosting — and each level buys different guarantees at very different operating cost. Choose the lowest level that meets guarantees you can name.

Hard cases exist: data that legally cannot leave, sovereignty mandates, air-gapped environments, extreme sustained volume. Everything else is usually served better by fixing the knowledge and permission layer than by hosting weights.

SOURCES AND FURTHER READING: Private LLM for Enterprise: When It Is Worth It ↗ · Private LLMs & Enterprise Knowledge hub ↗ · Service: AI Architecture Review ↗

Four layers, run on a calendar: a named owner per system, risk classes proportionate to consequence, controls embedded in the workflow rather than in a policy document, and evidence — evaluations, logs, decisions — you could show a customer or regulator.

The NIST AI Risk Management Framework supplies the operating verbs (govern, map, measure, manage) and the EU AI Act the legal risk taxonomy; borrow both rather than inventing structure.

SOURCES AND FURTHER READING: An AI Governance Framework CTOs Can Actually Run ↗ · NIST AI Risk Management Framework ↗ · Service: AI Readiness Assessment ↗

Readiness is always readiness for a specific capability, not a company-wide grade. Test one concrete use case across five dimensions — strategy, knowledge, people, operations, governance. Any question that produces silence in the room is a project in itself, found cheaply.

The two foundations that pay back regardless of use case: an executive owner whose objectives include the outcome, and a knowledge estate that is current, owned and permissioned.

SOURCES AND FURTHER READING: The AI Readiness Assessment: 20 Questions ↗ · Enterprise AI Strategy hub ↗ · Service: AI Readiness Assessment ↗

AI is owned wherever three things meet: the value case, the governance and the adoption. If those are three different people with no shared forum, the org chart label will not save you; if they are one accountable person, the title barely matters.

A dedicated Chief AI Officer earns its place when AI outcomes are material enough for the board agenda — with budget and authority attached, not as a signalling title. At Webskitters we chose the combined route: CTO to Chief Technology & AI Officer.

SOURCES AND FURTHER READING: CTO vs CIO vs Chief AI Officer ↗ · From CTO to CTAIO ↗

The Model Context Protocol is an open standard for connecting AI systems to tools and data sources — a common plug instead of one-off integrations. For enterprises it matters because agent capability becomes composable and auditable: you define what a tool exposes once, and any compliant model can use it under your controls.

The practical wins are permission clarity and vendor mobility: the tool boundary is where you enforce what an agent may do, and it survives model changes.

SOURCES AND FURTHER READING: MCP for Enterprise Commerce Agents: A Practical Playbook ↗ · Model Context Protocol ↗

The bottleneck moves. When agents draft more of the code, the constraint becomes verification: review, CI, integration and the judgment to accept or reject. Teams that redesign those loops — sharded builds, test-impact analysis, merge queues — convert agent speed into shipped software; teams that do not just queue faster.

The cultural change matters as much: ownership of an outcome, not authorship of every line, becomes the unit of accountability.

SOURCES AND FURTHER READING: Reworking CI for AI Teams with Multi-Agent Systems ↗ · What AI Changes About Engineering Culture ↗

Buy the layers that are becoming commodities — models, inference, generic tooling — and build the layers that encode your advantage: knowledge pipelines, workflow integration, evaluation on your own tasks. The model is the most replaceable part of the stack; your integration of it is not.

The build-or-buy line moves every year, so keep exits open: abstractions over model calls and evaluations that transfer between vendors.

SOURCES AND FURTHER READING: AI for Founders: A Complete Guide ↗ · Private LLM for Enterprise ↗ · Service: AI Strategy Advisory ↗

Never from leaderboards. Build a small benchmark suite from your own prompts, tools and latency constraints — fifty to two hundred cases with acceptance criteria you can score — and run every candidate through it, repeatedly, at honest volumes with real cost accounting.

The suite is a durable asset: every future model decision gets cheaper, and migration stops being a leap of faith.

SOURCES AND FURTHER READING: Claude Opus 5.5: A Decision Brief ↗ · The Enterprise AI Stack Explained ↗

It raises the value of judgment and lowers the value of typing. Review, architecture sense, taste and the willingness to say “this is wrong” matter more when generation is cheap. Leaders should protect ownership and honest evaluation, and resist measuring productivity by volume of output.

Psychological safety becomes a quality mechanism, not a comfort programme: teams that surface doubt early ship better systems.

SOURCES AND FURTHER READING: What AI Changes About Engineering Culture ↗ · Psychological Safety Is About Decision Quality ↗

Start with a workflow you already understand deeply, where time or cost visibly leaks, and apply assisted intelligence there — not with a moonshot product pivot. Measure against the old baseline, keep the human in the loop where judgment matters, and only scale what survives honest evaluation.

The founder advantage is speed of iteration, not size of budget. Use it on adoption loops, not on infrastructure you can rent.

SOURCES AND FURTHER READING: AI for Founders: A Complete Guide ↗ · Welcome to 2026 ↗