Custom AI assistant or Microsoft Copilot — which one should your business actually pay for? It is the most common AI buying question we hear from UK firms in 2026, and it is usually asked the wrong way round. The two options are not rivals for the same job. One is a licensed feature that lives inside Office apps you already own; the other is a piece of software built around your data, your workflows and your customers. Choosing well means understanding what each one genuinely does, what each one costs over three years, and where each one breaks.
This guide puts real UK figures on both sides. We have already priced Copilot in detail in our Microsoft 365 Copilot cost and ROI analysis and priced bespoke builds in our AI consulting cost UK guide; here we bring the two together into one decision. You will see a head-to-head comparison, a worked example that runs a 50-person firm down both routes, and a five-question framework that settles the choice for most businesses in an afternoon.
The short version: Copilot is the right first move for most companies, a custom AI assistant is the right move for a specific, valuable job Copilot cannot touch — and the strongest setups we deploy use both. The rest of this article earns those three claims with numbers.
Table of contents
- Microsoft Copilot and the Custom AI Assistant, Defined
- What Microsoft Copilot Does Well
- What a Custom AI Assistant Can Do That Copilot Cannot
- Microsoft Copilot vs Custom AI Assistant: The Head-to-Head
- What Each Route Costs a UK Business
- Worked Example: One Firm, Both Routes
- When Microsoft Copilot Is the Right Answer
- When a Custom AI Assistant Is the Right Answer
- The Hybrid Route: Copilot Plus a Custom AI Assistant
- Security, Data Protection and Governance
- How to Choose: A Five-Question Decision Framework
- Getting a Custom AI Assistant Built Without Wasting Money
- FAQ: Microsoft Copilot vs Custom AI Assistant
- References
Microsoft Copilot and the Custom AI Assistant, Defined
The comparison only makes sense once the two terms are pinned down, because vendors stretch both of them. Here is what each one means in practice for a UK business.
What Microsoft Copilot is
Microsoft 365 Copilot is a per-user licence that embeds a general-purpose assistant into Word, Excel, Outlook, Teams, PowerPoint and the Microsoft 365 web apps. It drafts, summarises, searches and answers using the documents, mail and meetings your tenant already holds, subject to each user’s existing permissions. You do not host anything, you do not choose the model, and you cannot change how it behaves beyond admin settings and lightweight agent tooling — we covered that layer separately in our Microsoft Copilot agents guide.
What a custom AI assistant is
A custom AI assistant is software your business owns or commissions: a language model — bought through an API or self-hosted — wrapped in your own retrieval layer, your own rules, your own interface and your own integrations. It can sit on your website answering customer questions, inside your line-of-business system chasing invoices, or on your intranet answering from twenty years of project files. The defining feature of a custom AI assistant is control: you decide what it knows, what it refuses, where the data flows and what it plugs into.
Why the two keep getting confused
Both are called “AI assistants”, both chat, and both demo beautifully. The confusion costs money in two directions. Firms buy Copilot expecting a bespoke tool and are disappointed when it cannot follow their claims process; other firms commission a custom AI assistant for £60,000 to do drafting and summarising that a £13.80 licence already does well. This article exists to stop both mistakes.
What Microsoft Copilot Does Well
Copilot’s strengths are exactly what you would expect from a product welded into the world’s dominant office suite. Played to those strengths, it is extremely good value.
Work inside documents, mail and meetings
Drafting and reworking documents, summarising long email threads, recapping Teams meetings, first-pass Excel analysis, turning notes into slides — this is Copilot’s home turf. The measured evidence is genuinely decent. The UK Government Digital Service ran a cross-government trial with over 20,000 staff and found an average saving of 26 minutes per day; a separate DWP evaluation of 3,549 staff measured 19 minutes per day. Honesty demands the third figure too: the Department for Business and Trade found no discernible productivity gain in its own evaluation. The tool is real, and so is the variance.
Copilot works best when the value is spread thinly across everyone — a few minutes here and there, across every role that lives in Outlook and Word all day.
Deployment speed and low commitment
Copilot deploys in days: assign licences, set the admin controls, train people. There is no project risk, no code to maintain, and you can cancel at the next renewal. For a business that wants its first structured AI win this quarter, that speed is worth a great deal — provided the tenant is tidy, which is exactly what our Copilot readiness assessment checklist tests.
Governance you inherit rather than build
Copilot honours the permissions, retention and audit arrangements already in your Microsoft 365 tenant. Prompts and responses stay within Microsoft’s service boundary and are not used to train foundation models. You still have governance work to do — oversharing in SharePoint becomes very visible, very fast — but you are configuring controls, not designing them.
What a Custom AI Assistant Can Do That Copilot Cannot
A custom AI assistant justifies its price only by doing things the licence physically cannot. There are more of those than Copilot’s marketing suggests.
Serve your customers, not just your staff
Copilot is internal by design: it assists the licensed user. It cannot sit on your website at 2am answering product questions, triaging support tickets in your brand voice, or qualifying leads. A customer-facing custom AI assistant is a different product category entirely, and it is one of the few places where assistant projects generate revenue rather than saved minutes.
Follow your process, not a generic one
A custom AI assistant can encode the way your business actually works: check the certificate register before answering, quote only from the current price list, refuse anything outside policy, escalate edge cases to a named person, write the result back into the CRM. It is a worker in a workflow — closer to the AI employees and autonomous agents we build for clients than to an autocomplete. Copilot’s agent tooling gestures at this, but consumption pricing and platform limits arrive quickly; we priced those limits in the agents guide above.
Answer from any data, wherever it lives
Copilot grounds itself in Microsoft 365. The knowledge your business runs on often is not there: the job-costing database, the practice management system, the SQL warehouse, the supplier portal, the archive of scanned PDFs. A custom AI assistant with a retrieval pipeline built for natural language processing can index all of it — and answer with citations back to source, under access rules you define, as we set out in our private AI for UK businesses guide.
Control the model, the residency and the exit
With a custom AI assistant you choose the model and can change it — swap to a cheaper one when quality allows, a stronger one when it matters, or a self-hosted open-weight model when data must never leave the building. Firms with strict confidentiality duties can even train LLMs on their own data. None of those levers exists with a licence: Copilot’s model, residency options and roadmap are Microsoft’s decisions, not yours.
Microsoft Copilot vs Custom AI Assistant: The Head-to-Head
The table below is the whole argument in one place. Neither column wins every row — which is precisely why the decision framework later in this article matters more than any single line.
| Factor | Microsoft Copilot | Custom AI assistant |
|---|---|---|
| Upfront cost | None | Typically £25,000–£120,000 to build |
| Ongoing cost shape | Fixed per user per month | Usage-based inference plus hosting and upkeep |
| Time to live | Days to weeks | Two to five months |
| Where it works | Inside Microsoft 365 apps, for licensed staff | Anywhere — website, systems, customers, staff |
| Knows your data | Tenant content only, per-user permissions | Any source you connect, any rules you set |
| Customisation depth | Settings and lightweight agents | Full control of model, behaviour and interface |
| Maintenance burden | Microsoft’s problem | Yours — monitoring, evaluation, updates |
| Project risk | Near zero | Real — scope, data quality, adoption |
| Exit | Cancel at renewal, keep nothing | You own the software and the pipeline |
The maintenance row deserves emphasis. A licence never rots; software does. Any custom AI assistant needs monitoring, periodic re-evaluation against new models, and someone accountable for it — budget roughly a third to a half of the build cost again per year, a ratio we have verified across real projects.
What Each Route Costs a UK Business
Prices below are the UK figures we verified from Microsoft’s own price list in August 2026 (annual commitment, excluding VAT) alongside the build economics from our consulting cost research.
| Route | What you pay | Notes |
|---|---|---|
| Copilot add-on licence | £13.80 per user/month | Promotional to 30 Sep 2026; list £16.10 |
| Business Standard with Copilot | £18.10 per user/month | £7.30 uplift on the £10.80 base plan |
| Business Premium with Copilot | £24.60 per user/month | £7.70 uplift on the £16.90 base plan |
| Copilot Credits (agents) | £153.80 per 25,000-credit pack/month | Agent actions cost 5 credits; grounding 10 |
| Custom build (API-based) | £25,000–£120,000 once, then usage | Scope-dependent; consultancy near £1,100/day |
| Custom, self-hosted model | Build cost plus hardware or GPU rental | For data that must not leave your estate |
The running-cost comparison at small scale
For a 20-person team, our private AI cost modelling put the monthly figures side by side: the Copilot add-on comes to £276 (20 × £13.80), an API-based assistant handling the same team’s queries runs about £108 in metered usage, a rented GPU server for a self-hosted model about £264, and an owned inference machine about £155 once hardware is amortised. Ongoing cost is genuinely competitive for the custom routes — it is the build cost and the upkeep that separate them.
Why custom quotes vary so much
Two builds described with the same sentence can differ by £80,000. The drivers are data readiness (cleaning and structuring the knowledge sources is routinely the largest line), integration count, evaluation depth, and whether the assistant merely answers or actually acts. Day rates are the stable part — contractor medians sit near £575–£588 and consultancy rates near £1,100 — so the variance is almost entirely scope.
Worked Example: One Firm, Both Routes
Take a 50-person UK professional services firm. It wants staff to stop losing time to email and documents, and it wants client questions answered from its own knowledge base rather than from memory. Here is the honest arithmetic for each route over three years.
Route one: Copilot for everyone
Fifty add-on licences at £13.80 is £690 a month — £8,280 a year, £24,840 over three years. Rollout takes a fortnight plus training. Using the government trial range of 19–26 minutes a day, the licence pays for itself if each person genuinely banks even a fraction of that saving.
Route two: a custom AI assistant on the firm’s knowledge
A retrieval-based custom AI assistant over the firm’s documents and systems prices out at 45 consultancy days — £49,500 at £1,100 a day. Running it: roughly 40,000 queries a month at ~3,000 input tokens each is 120 million input tokens (£240 at £2 per million) plus 16 million output tokens (£128 at £8 per million), call it £368 of inference; hosting and the vector database about £120; monitoring and maintenance retainer about £800. That is ~£1,288 a month, £15,456 a year — £95,868 over three years including the build.
| Build phase | Days | Cost at £1,100/day |
|---|---|---|
| Discovery and scoping | 5 | £5,500 |
| Data preparation and indexing | 10 | £11,000 |
| Assistant build and prompt design | 12 | £13,200 |
| Evaluation and testing | 5 | £5,500 |
| Integration with firm systems | 5 | £5,500 |
| Security review | 3 | £3,300 |
| Hypercare and handover | 5 | £5,500 |
| Total build | 45 | £49,500 |
What the gap has to earn
The custom route costs £71,028 more over three years. Spread across 50 people and 156 weeks, it must deliver about £9.11 per person per week of extra value beyond what Copilot would have delivered — roughly 16 minutes a week at £34 an hour. That is a clearable bar for an assistant that genuinely removes lookup work or answers client queries directly, and an uncleared one for an assistant that merely duplicates drafting Copilot already does. This calculation, run honestly on your own numbers, is the whole decision.
When Microsoft Copilot Is the Right Answer
Copilot wins whenever the value you are chasing is diffuse, internal and document-shaped. In practice that covers most firms’ first year of AI adoption.
The clear Copilot cases
Buy the licence when the pain is email volume, meeting load, document drafting and first-pass analysis; when your data already lives in Microsoft 365; when you have no appetite for a software project; or when you need something live this month. Buy it also as a learning instrument — six months of Copilot usage data will tell you more about where a custom AI assistant would pay than any workshop.
The honest caveats
Licence maths punishes optimism: seats assigned to people who never open the tool are pure cost, and the DBT result above shows adoption is not automatic. Roll it out to the teams whose week is genuinely made of documents, measure, then widen. And remember the ONS finding that AI use among UK businesses has roughly tripled since 2023 to about 35% — your competitors’ baseline is rising whether you buy or not.
When a Custom AI Assistant Is the Right Answer
A custom AI assistant wins when the job is specific, high-value and outside Copilot’s walls. The pattern across successful projects is a named process with a number attached.
The clear custom cases
Commission a custom AI assistant when customers need answering, not just staff; when the knowledge lives outside Microsoft 365 in databases, legacy systems or archives; when the assistant must act — create records, chase debtors, produce quotes — under rules you define; when confidentiality or residency demands a private deployment; or when per-seat pricing collapses at your scale, because a thousand light users cost £13,800 a month on licences but a few hundred pounds through an API.
The honest caveats
The failure statistics are brutal and worth staring at. MIT’s 2025 State of AI in Business research found 95% of organisations saw no measurable return on roughly $30–40 billion of generative AI spending, with only about 5% of custom tools reaching production; Gartner projects that over 40% of agentic AI projects will be cancelled by the end of 2027. The survivors share three habits: one narrow use case, data cleaned before the build, and a maintenance budget from day one. A custom AI assistant is software — treat it with the discipline you would give any other build, starting with an AI readiness assessment before any code is written.
The Hybrid Route: Copilot Plus a Custom AI Assistant
The framing of Copilot versus custom is, for many firms, a false choice. The strongest pattern we deploy in 2026 is both, with a clean division of labour.
How the split works
Copilot covers the horizontal layer: every member of staff, general document and email work, a fixed £13.80 a seat. The custom AI assistant covers one vertical, valuable process: the client-facing answer desk, the tender-response engine, the claims triage bot. Neither duplicates the other, so neither’s cost is wasted — the licence handles breadth, the build handles depth.
Why the sequence matters
Do Copilot first. It is live in weeks, it normalises AI-assisted work, and its usage patterns reveal exactly where people burn time — which is the discovery work for the custom build, done free. Firms that start with the custom AI assistant too often spend the build budget on a use case Copilot would have covered, then discover the real bottleneck was somewhere else entirely.
Security, Data Protection and Governance
Both routes can be run safely in the UK; they distribute the work differently. With Copilot you inherit controls; with a custom AI assistant you must specify them.
The rules common to both
If prompts or retrieved documents contain information about identifiable people, UK GDPR applies and the ICO’s guidance on AI and data protection expects lawful basis, transparency and minimisation to be settled before deployment. Confidentiality duties reach further than data protection — client papers pasted into any tool you have no contract with can breach an NDA with no personal data involved.
Where each route carries risk
Copilot’s characteristic failure is oversharing: it reads everything each user can read, so stale permissions become search results. Fix SharePoint hygiene before rollout. A custom AI assistant’s characteristic failures are the OWASP LLM risks — prompt injection through retrieved content above all — plus everything ordinary software inherits: authentication, logging, dependency updates. Build against the NCSC’s guidelines for secure AI system development, evaluate against the NIST AI Risk Management Framework, and give the assistant an owner with a maintenance budget, not just a launch date.
How to Choose: A Five-Question Decision Framework
Answer these five questions in order. Most UK businesses reach a defensible decision without a consultant in the room.
The five questions
First — who is served? Staff only points to Copilot; customers or external users point to a custom AI assistant. Second — where does the knowledge live? Microsoft 365 favours the licence; databases and legacy systems favour the build. Third — does it need to act, or only assist? Actions under business rules need custom control. Fourth — what does the gap have to earn? Run the worked-example arithmetic on your own headcount. Fifth — can you feed a build? No clean data, no owner, no budget for upkeep means buy the licence and revisit in a year.
| Your situation | Better fit | Why |
|---|---|---|
| Email, documents and meetings dominate the week | Copilot | Its home ground, measured savings, zero project risk |
| Customers need answers around the clock | Custom | Copilot cannot face customers at all |
| Knowledge sits in line-of-business systems | Custom | Copilot grounds only in Microsoft 365 |
| First structured AI adoption, live this quarter | Copilot | Days to deploy, cancel at renewal |
| One process is worth six figures a year | Custom | Depth beats breadth where the value is concentrated |
| Strict residency or confidentiality duties | Custom (private) | Self-hosted deployment keeps data in your estate |
| Hundreds of light users, occasional queries | Custom (API) | Usage pricing undercuts per-seat licences at scale |
| Document work now, one valuable process later | Both, Copilot first | Licence for breadth, build for depth, usage data guides the build |
Getting a Custom AI Assistant Built Without Wasting Money
If the framework points you at a build, the difference between the 5% that reach production and the 95% that do not is mostly procurement discipline. Four rules cover it.
Scope one job, prove it, then extend
Commission the custom AI assistant to do one thing with a measurable number attached — answer client queries with 90% deflection, cut tender first-drafts from two days to two hours. Resist every “while we’re at it”. Extensions are cheap after a working v1 and ruinous before it.
Spend on data before models
The model is the commodity; your data pipeline is the product. Expect data preparation to be the biggest phase of the build, as it was in the worked example above, and treat any supplier who skips discovery and quotes a fixed price on a phone call as a red flag.
Demand evaluation, not demos
A demo proves the happy path. Insist on a written evaluation set — real questions, marked answers, refusal cases — and a measured score before go-live, then re-run it monthly. This is also your protection when models change underneath the assistant.
Contract for exit and upkeep
Own the code, the prompts, the index and the evaluation set. Price the maintenance retainer into the business case at a third to half of build cost per year. And keep the assistant’s workflow automation integrations documented, so the next supplier can take it over without archaeology.
FAQ: Microsoft Copilot vs Custom AI Assistant
Is Microsoft Copilot cheaper than a custom AI assistant?
Almost always, yes. In our worked example the three-year cost was £24,840 for 50 Copilot seats against £95,868 for the build — the custom AI assistant costs roughly four times as much. The exception is scale with light usage: very large user counts on occasional queries can make API-metered custom cheaper than per-seat licensing.
Can a custom AI assistant read my Microsoft 365 data like Copilot does?
Yes — the Microsoft Graph exposes mail, files and calendars to properly authorised applications, so a custom AI assistant can ground itself in tenant content and other sources at once. The build must respect per-user permissions, which Copilot gives you for free; budget integration days for it.
How long does a custom AI assistant take to build?
A single-purpose retrieval assistant typically takes two to five months from discovery to production — 45 build days in our worked example — with data preparation the phase most likely to stretch. Anything quoted in a fortnight is a demo, not a deployment.
Do I still need Copilot if I commission a custom AI assistant?
Usually yes, and the hybrid section above is why: the licence covers everyone’s document work for £13.80 a seat, which no sane custom build should replicate. The two overlap far less than their marketing suggests.
What happens to a custom AI assistant when better models arrive?
A well-built one improves for pennies: swap the API model, re-run the evaluation set, ship. This is the strongest argument for owning the retrieval layer and evaluation set — the parts that appreciate — while renting the model, the part that depreciates.
References
Microsoft 365 Copilot overview — Microsoft Learn
Microsoft 365 Business Standard UK pricing — Microsoft
Microsoft 365 Copilot Experiment: Cross-Government Findings Report — GOV.UK
Guidelines for Secure AI System Development — NCSC
Artificial Intelligence Guidance — Information Commissioner’s Office
AI Risk Management Framework — NIST AIRC
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.