Answers
The questions people ask before hiring an AI consultancy — answered straight.
Real numbers, real trade-offs, and the projects we would tell you not to build. No gated PDFs.
The AI glossary for the person signing the contract
39 terms from AI proposals, defined plainly — then what each one changes about a buying decision.
Buying AI services
How much does an AI consultant cost?
In 2026, independent AI consultants charge roughly $120–$400 per hour, while AI consultancies price by engagement: a fixed-scope AI opportunity assessment runs $999–$15,000, a 2–6 week strategy or proof-of-concept sprint runs $15,000–$60,000, and an embedded engineering retainer runs $12,000–$45,000 per month per senior engineer pod. Offshore and hybrid teams sit at the lower end of each band without a quality penalty when the engineers are senior. The cheapest useful starting point is a paid fixed-scope assessment, because it converts an open-ended question into a costed plan before you commit to a build.
Read the answer→How do I choose an AI consultancy?
Choose an AI consultancy by demanding three things before you sign: a production system they built that is still running, a named business metric it moved, and the actual engineers who will do your work on the first call. Start with a small paid fixed-scope engagement — an assessment or a two-week proof of concept on your own data — rather than a long discovery phase, so you buy evidence instead of promises. The strongest predictor of failure is a firm that answers technical questions with case studies and staffs the build with people you never met.
Read the answer→How long does it take to build an AI system?
A focused AI system reaches production in 6–12 weeks: roughly 1–2 weeks to scope and assess, 2–3 weeks to a working prototype on your real data, and 3–7 weeks to harden it into something that runs unattended. Timelines slip past this almost entirely for non-technical reasons — data access approvals, security review, and no single decision-maker — rather than because the model work is hard. Treat any proposal quoting under four weeks to production as skipping evaluation, and anything quoting over six months as a programme rather than a project.
Read the answer→How do we measure ROI on an AI project?
Name one business metric per use case, record its baseline before any build starts, and write the payback arithmetic down while it can still change the decision. Measure against a holdout — a comparable team, queue or period still working the old way — because a before-and-after comparison absorbs every other change your business made that quarter. Count the full cost, which means build, inference at real volume, and the ongoing maintenance an AI system needs to resist drift, not just the invoice. In back-office automation a use case that cannot show payback within twelve months usually has a scope problem rather than a technology problem.
Read the answer→Should we build AI ourselves or buy an off-the-shelf tool?
Buy when the workflow is standard across your industry — transcription, meeting notes, generic support deflection, code assistance — because a vendor amortises that build across thousands of customers and you will not beat the economics. Build when the process is specific to how your company competes, or when the data lives in systems no vendor integrates with, which is the usual reason mid-market automation has no off-the-shelf answer. The deciding question is not cost: it is whether the unusual part of the problem is your data and workflow or the AI itself, and if it is the AI, buy. Most companies end up doing both, and the expensive mistakes are building a commodity and buying a differentiator.
Read the answer→What is a forward deployed engineer?
A forward deployed engineer is a software engineer who embeds with a customer's team and builds production software inside that customer's environment — their repository, their data, their access controls — rather than delivering a general product or a written recommendation. The term was popularised by Palantir and has become the common label for how applied AI work is delivered, because AI systems depend so heavily on a specific company's data and processes that a generic build rarely survives contact with them. The practical test is who holds the keyboard: an FDE engagement leaves running code you own, an advisory engagement leaves a document about code somebody else will write. It is normally priced as a monthly retainer per engineer or per small pod rather than per deliverable.
Read the answer→Funded startups & product teams
RAG or fine-tuning — which should we use?
Use RAG when the model lacks knowledge, and fine-tuning when it lacks behaviour. If your problem is that the model does not know your documents, policies or product data, retrieval fixes it; if the model knows enough but answers in the wrong format, tone or reasoning style, fine-tuning fixes it. For roughly 80% of business use cases the correct answer in 2026 is a strong prompt plus RAG plus a proper evaluation set, because it ships in weeks, updates the moment your data changes, and costs far less than maintaining a fine-tuned model.
Read the answer→Should our startup build AI features in-house or hire help?
A funded startup should bring in an embedded AI partner for the first production AI feature, and hire in-house only once that feature is proven and load-bearing. Hiring a senior AI engineer takes 3–6 months and roughly $200,000 a year fully loaded, before you know which capability you need permanently — whereas an embedded partner ships the first version in 6–12 weeks and leaves your team owning the code. The rule of thumb: outsource discovery and the first build, insource whatever becomes core to your product.
Read the answer→Do AI agents actually work in production?
AI agents work reliably in production when they are narrow, reversible and supervised — research and summarisation, code and content drafting, ticket triage and routing, data extraction across systems, and internal operations with a human approving the final action. They fail when given open-ended goals, irreversible actions, or long chains of steps where a single early error compounds silently. The useful design question in 2026 is not how autonomous an agent can be, but how cheaply a wrong action can be caught and undone.
Read the answer→Mid-market operations
Where does AI actually save money in a mid-market company?
In mid-market companies, AI pays back fastest on high-volume, low-stakes internal work: document intake and data extraction, support ticket triage and draft replies, quote and proposal generation, invoice and claims matching, and internal knowledge search. These typically return their build cost in 4–9 months because they replace hours of repetitive reading and re-keying rather than attempting judgement. The projects that consistently disappoint are customer-facing autonomous agents, company-wide 'AI transformation' programmes, and anything that requires clean data you do not yet have.
Read the answer→Is our data safe with an AI consultancy?
Your data is as safe as your contract and architecture make it, and both are entirely within your control. Require four things before any access is granted: a written no-training clause covering the consultancy and every model provider in the chain, named data residency with a Data Processing Agreement where personal data is in scope, least-privilege access to redacted or sampled data rather than production, and full IP assignment of code, prompts and evaluation sets. Enterprise API tiers from the major model providers do not train on submitted data by default — the real exposure is usually an over-permissioned integration or an unlogged internal tool, not the model provider.
Read the answer→Running AI in production
Which AI model should we use?
Choose by task and constraint rather than by leaderboard position. Prove the use case works on a frontier model first, then move down the price ladder to the cheapest model that still passes your own evaluation set — for classification, routing and extraction that is usually a small model, and for long multi-step reasoning it usually is not. The decision that actually matters is not which vendor you pick but whether you built an evaluation set before picking, because without one every comparison is a matter of taste. Keep the model behind an abstraction so switching is a configuration change, since the ranking will move again within months.
Read the answer→How do we cut our AI running costs?
Measure cost per transaction before changing anything, because a monthly total tells you nothing about which feature is expensive. In most production systems the majority of spend is input tokens — retrieved context and long system prompts — rather than the model's output, so the highest-return fixes are trimming what you send, caching the stable prefix of every prompt, and routing easy requests to a smaller model. Those three commonly halve a bill without touching answer quality, and all of them are cheaper than renegotiating rates. Do not optimise a feature whose behaviour is still changing weekly: distillation and model downgrades cost more engineering time than they save until the scope has settled.
Read the answer→Next step
Still deciding? Get the answer for your business, not the general case.
The AI Opportunity Assessment applies all of the above to your actual processes, data and numbers — and tells you plainly if the answer is 'not yet'.