AI adoption ยท Platforms
Choose AI tools around the work
Pick the tool from the work rather than the other way round. What the work looks like, what you already licence, which systems the tool has to reach, who has to check the output and how many people will genuinely use it will settle the shortlist between Microsoft 365 Copilot, Claude, ChatGPT, Google and a custom build. A combination of them is a legitimate answer rather than a sign of indecision.
What follows is the criteria and the routes worth investigating, so you can narrow the list yourself before anyone quotes you anything.
What decides it
What actually decides it?
Seven criteria. Work through them before looking at any product page. The first three are the ones that tend to narrow a shortlist, and they do it without anyone having to read a feature table.
- The shape of the work
- Long documents and analysis pull one way. Short in-application tasks across email, slides and spreadsheets pull another. High-volume rule-shaped processing pulls towards a build rather than a seat.
- What you already run
- A Microsoft-resident company starts with a narrower question than one on Google Workspace. Existing spend is sunk, so only the incremental cost belongs in the comparison.
- Sources and connections
- Which systems does the tool have to read, and write? A connector that does not exist for your case system is a build, whatever the brochure implies.
- Controls
- Admin, retention, logging, what leaves your tenant, and what your clients or regulator require in writing. These are per-product and per-region answers, and they are checked against vendor documentation on the day, not assumed.
- Who is meant to use it
- Everyone, or twenty people whose work is depth? The second costs less and changes more. Buying seats for a population that will not open them is the commonest way to lose the business case.
- Review requirements
- Output somebody has to check is not free. If every result carries a name on it, the reviewer time belongs in the cost side and it changes the ranking.
- The whole operating cost
- Licences, consumption, integration and data preparation, training time, review and ongoing support. Check what consumption charges apply to the specific plans on your shortlist, because a charge that sits under the licence line is the one most often left out of the sum.
The five routes
Which route should we investigate?
Five routes, each with a starting situation that makes it worth a look. These are starting points for investigation rather than recommendations, and more than one row can apply to you at once.
| If this is your starting situation | The route worth investigating | What discovery has to check | Where to read more |
|---|---|---|---|
| Daily work already sits in Microsoft 365 | Microsoft 365 Copilot, inside the applications people already have open | Which licence actually applies to you, whether SharePoint and OneDrive permissions are correct before anything reads them, source quality, and where configuration work is needed beyond switching it on | The Copilot playbook |
| Document-heavy work with methods worth reusing | Claude, where the value is long documents, analysis and encoding how your company does things | Plan and feature fit, which sources it needs to reach, governance and review, and where assistant use stops and a build starts | The Claude implementation guide |
| Broad research, analysis and drafting across teams | ChatGPT, as a general assistant with wide familiarity already in the building | Workspace controls and admin, which connections exist for your systems, plan eligibility, and which workflows would need separate API work rather than the product | The ChatGPT playbook |
| Google Workspace is where the work lives, or there is a Google Cloud estate | Gemini in Workspace for in-suite assistance, or Gemini Enterprise on Google Cloud for a built solution. These are separate products and separate decisions | Which of the two is actually being proposed, the deployment route, the admin controls, and how it would reach anything that still lives in Microsoft 365 | Gemini Enterprise implementation |
| A workflow needs triggers, writes into core systems or continuous running | Custom implementation, in your own cloud, using whichever models suit each step | Whether a configured product already covers it, and the honest cost of integration, evaluation, human review and keeping it running | AI agents and custom builds |
For a feature-by-feature comparison, the detailed head-to-head sits in the comparison article, and the general AI-native guide works through eighteen tasks across the same routes.
Mixed estates
Is using two platforms a mistake?
No. One shape worth considering in a Microsoft-resident company is a governed assistant inside the applications everyone already uses, a second tool for the smaller group whose work is depth, and a custom build for a process that neither seat product finishes. Two paired guides work through that decision, for Copilot alongside Claude and for Copilot alongside ChatGPT.
What turns that from sensible into expensive is buying every tool for every person. Count a shared platform cost once across the use cases that ride on it, show the overlap where the finance director can see it, and size each tool to the group that will actually open it weekly.
Before you commit
What does the pilot have to prove?
A shortlist is a hypothesis. The pilot is what settles it, and it needs all four of these rather than three.
- Quality on your own workMeasured on the same sample as the baseline, including exceptions, and paired with a quality check rather than a speed figure on its own.
- Weekly use without chasingBy the people it was bought for. This is the number the business case turns on, so measure it rather than assume it.
- Controls in place and evidencedAccess, retention, logging and whatever your clients or regulator require, demonstrated rather than asserted.
- Cost matching the assumptionIncluding the metered consumption line and the reviewer time, against what the case assumed before it started.
Common questions about choosing an AI platform
Which AI platform is best?
There is no answer to that question without the work in front of you. The tool follows from what the work looks like, what you already licence, which systems it has to reach, who has to review the output and how many people will genuinely use it. A company doing long document analysis and a company doing short tasks inside Outlook and Excel will reach different shortlists from the same budget, and both will be right.
We already pay for Microsoft 365. Does that settle it?
It narrows it and does not settle it. Copilot is the route to evaluate first when the work already happens inside Microsoft 365, because it starts from licences you may already hold. Whether it is the right answer for every group depends on what each group actually does, which sources the assistant has to reach, and the whole cost including anything metered. The group whose day is long documents and the workflow that has to start on an event and write into a core system are both worth testing separately rather than assuming either way. Whatever you choose, check the permissions on the content it would be able to read before you switch it on.
Can we use more than one platform?
Yes. One shape worth considering is a broad assistant for every seat, a second tool for the people whose work is depth, and a custom build for a process neither finishes. What turns that from sensible into expensive is buying all three for everybody. Count a shared platform cost once across the use cases that ride on it, and put the overlap on the page where the finance director can see it.
Why are there no prices on this page?
Because plans, model line-ups and the consumption charges underneath them move often enough that a figure printed here would be out of date by the time you acted on it. For a real number, two things are worth doing. Check the current published position against vendor documentation for the specific plans on your shortlist, which we will do with you. Then put it into the business case calculator alongside integration, training, review and support, because the licence line on its own will not tell you what the year costs.
What about data residency?
It is a real question and it is product and region specific rather than a property of a vendor. Some products keep processing inside a boundary you already accept, some store in one region and process in another, and some depend on which model you have enabled. For most companies the answer is documentation done properly rather than a different vendor, and it only becomes a hard constraint where a client mandate, a public-sector contract or the kind of data involved makes it one. We set out what applies to the products you are actually considering.
What does a pilot have to prove before we commit?
Four things. That the output is good enough on your own work, measured on the same sample as the baseline. That the people it was bought for open it weekly without being chased. That the controls your company needs are in place and evidenced. And that the cost, including the metered lines and the review time, matches what the business case assumed. A pilot that proves three of the four has told you something useful and has not told you to buy.
Bring the work, and we will narrow the list
Tell us what the work looks like and what you already run. We will say which routes are worth investigating and what would have to be true for each one.
The first call covers what you want to improve, the tools you already run, and whether a scoped paid discovery is the right next step.