Before You Buy the Platform: A Build-vs-Buy Worksheet for Mid-Market AI
Vendors will sell you a platform. The decision is whether you need one.
The pitch deck moment
A vendor walks into a 400-person European company with a deck. The deck has a unified copilot, a connector library, a governance layer, and an enterprise plan. The math at $30 per user per month for Microsoft 365 Copilot lands at roughly $144,000 a year for that headcount, before the implementation partner adds a number. The CIO leaves the meeting unsure whether to feel sold-to or relieved.
Most companies in this position do not need a platform. They need three workflows shipped. The trouble is that "three workflows shipped" does not arrive in a deck on Tuesday. So the platform wins the calendar.
This is a worksheet. It is what we use ourselves when a client asks whether to buy or to build. It will not tell you what to do. It will tell you what conversation to have inside before you have it with the vendor.
$30/user/moMicrosoft 365 Copilot enterprise pricing, list (Microsoft, 2026)What "platform" actually means in 2026
Three things hide under the word.
Category one is a workflow runtime. Examples: Inngest (free tier covers 50,000 runs per month, then usage-based), Temporal Cloud ($25 per million actions, list), self-hosted Temporal, n8n. Built to compose calls between systems. Cheap. Strong tooling. Requires engineering ownership.
Category two is an agent platform. Examples: LangGraph Cloud, CrewAI Enterprise, the AI gateways from the major clouds. Built to run agentic loops with retries and observability. Mid-priced. Adds a runtime concept your team has to learn.
Category three is a horizontal copilot. Examples: Microsoft 365 Copilot ($30/user/month enterprise; $21/user/month for Business under 300 seats), Glean, Sana, Notion AI. Built to give every employee a chat interface over your data. Most expensive at scale. Lowest engineering lift. Highest risk of becoming the chatbot nobody uses.
These are not the same product. Vendors will conflate them in conversation. Do not let them.
The worksheet
Five questions. Score each one to five.
1. How many distinct workflows do you actually want to ship in the next twelve months?
One or two: build. Three to ten: build, plus a workflow runtime. More than ten: a platform might pay back. McKinsey's 2025 survey shows that only 31% of organisations have managed to scale AI enterprise-wide, and the bottleneck for the other 69% is not picking the right platform. It is choosing fewer workflows and finishing them.
2. How specific is the work to your operation?
If the workflow encodes something you do that nobody else does (a claims process, a Nordic-language compliance check, a legacy-CAD search), buying does not help. The platform vendor has not seen your work. Build it.
If the workflow is generic (employee Q&A over policy documents, generic meeting summaries, sales pipeline notes), buying might be cheaper.
3. How long is your contracting cycle for a six-figure software purchase?
If the answer is six months or more, buying loses on speed. By the time the purchase order clears, your three workflows could have been live for two quarters. Gartner's 2025 forecast that 30% of GenAI projects will be abandoned after the POC phase has a sister statistic that is rarely cited: a meaningful share of the abandonment comes from contract cycles that outlast executive attention.
4. How much engineering capacity do you actually have, in headcount that ships?
Not headcount on the org chart. Headcount on a Friday. If the answer is fewer than two people, buying is the answer regardless of the other questions, because nobody will be there to maintain what you build.
5. What is your auditability bar?
If you are regulated (banking, insurance, public sector), ask the vendor to show you the audit trail of an agent decision from six months ago. Most platforms cannot. The build path produces logs you control. The buy path produces logs the vendor controls.
How to read the score
Add the five scores. Three to ten: build, you are not ready for a platform. Eleven to seventeen: workflow runtime plus build, this is where most mid-market sits. Eighteen to twenty-five: a platform purchase could pay back, but verify with the cost ranges below.
Indicative first-year costs (USD, 400-person company, list prices)
| Path | Software (list) | Engineering | Total Y1 | Steady state Y2 |
|---|---|---|---|---|
| Build on a workflow runtime (Inngest free tier or Temporal $25/M actions) | 0–6,000 | ~90,000 (0.7 FTE) | ~90,000 | ~70,000 |
| Build with an agent platform (LangGraph Cloud / similar) | ~30,000 | ~60,000 (0.5 FTE) | ~90,000 | ~50,000 |
| Buy a horizontal copilot (Microsoft 365 Copilot at 400 seats) | 144,000 | ~40,000 (0.3 FTE rollout) | ~184,000 | ~144,000 |
| Hybrid: copilot plus runtime | ~150,000 | ~80,000 | ~230,000 | ~170,000 |
Software figures are list prices from the vendor pricing pages cited at the bottom of this piece. Engineering estimates are based on what the engagements we have run actually take, and will vary with seniority, geography, and existing infrastructure.
The build paths look more expensive in years three and four if the workflows pile up. The buy path looks more expensive every year because the contract escalates and the seat count drifts.
The build vs buy answer is rarely "one or the other." Most European mid-market companies need two or three workflows shipped on a runtime they own, plus optionally a copilot for generic Q&A. They are sold a platform that bundles both at multiple times the cost.
When to buy without thinking
Three situations where buying is the right answer even if the worksheet says otherwise.
One: you have a regulated document estate (legal, medical, financial) where the cost of getting search wrong is higher than the cost of the contract. Buy a copilot with vendor-side compliance certifications.
Two: you have a global helpdesk that gets the same fifty questions every day in eight languages. Buy a copilot. Build the policies it answers from.
Three: your engineering function does not exist or is fully booked on revenue work. Buy. Then make a case for hiring.
The webinar
On May 8 we walk through the five patterns we see inside European companies that ship AI without buying a platform first. The build-vs-buy decision is the third pattern. If your procurement meeting is Friday, this is the session to attend Thursday.
Register for AI That Works at Your Company →
If you want a second pair of eyes on a specific vendor pitch, send the deck to [email protected]. We will read it and reply within two days with the two questions worth asking the vendor before you sign.
Microsoft 365 Copilot Pricing (enterprise + business)(2026) Inngest Pricing(2026) Temporal Cloud Pricing(2026) McKinsey State of AI 2025(2025) Gartner GenAI Project Forecasts(2025) Calmworks engagement experience (anonymised)(2024-2026)Design & Technology
Next read.
2 relatedBeyond the Countdown: An Operational Playbook for the EU AI Act
Most organisations now know the AI Act is real. Far fewer have translated the five required capabilities into a calendar with names. This is the operational playbook we use with clients who still have time to ship.
Your Tools Already Talk. You Just Can't Hear Them.
We keep being told enterprises need more data. They don't. They need the data they already have to move between the systems they already use, without a human in the middle doing the translation.
Want to take this further?
Calmworks is an intelligence-first agency. Book 30 minutes and we'll show you what we'd do with this in your context.