Build vs buy an AI agent: the arithmetic that decides it

You cannot answer it until you know how many billable actions one conversation takes, because every pricing model now meters actions rather than conversations. That makes your bill a consequence of how the agent is designed, and it means the same use case can land either side of the line depending on choices you have not made yet.

JarvisBitz

Build or buy is usually asked before the one number that decides it is known. Both of the obvious platforms, Microsoft Copilot Studio and Salesforce Agentforce, now meter consumption rather than conversations, and so does a custom build. The unit they meter is roughly the action: a lookup, a generated answer, a record update, a grounded retrieval. Until you know how many of those one conversation takes, none of the three has a price.

That is the useful reframing. Your monthly bill is not a property of the vendor you pick, it is a consequence of how the agent is designed, and the same use case lands on either side of the line depending on decisions nobody has made yet.

Both platforms meter actions, not conversations

Microsoft publishes a rate card. A classic answer costs 1 Copilot Credit, a generative answer 2, an agent action 5, and tenant graph grounding 10. Credits stack within a single turn, and Microsoft's own documentation gives the example: an agent grounded in a tenant graph "could use 12 Copilot Credits (10 Copilot Credits for tenant graph grounding, and 2 Copilot Credits for generative answers) to respond to a single complex prompt". Credits are bought at $200 per pack of 25,000, which is $0.008 each.

Salesforce meters the same way under Flex Credits, where "each Action, a specific function that an AI agent executes on the platform, such as updating a record, summarizing a complex case, answering a product inquiry, or executing a custom prompt or flow, draws from a pool of Flex Credits". A standard action costs 20 Flex Credits, a voice action 30, and credits are $500 per 100,000, which is $0.005 each. So one standard Agentforce action is 10 cents.

Salesforce also offers $2 per conversation for customer-facing agents, and the two models are mutually exclusive: Flex Credits and Conversations are not supported in the same org. That makes the choice between them a straightforward calculation. At 10 cents per action, $2 buys 20 actions, so under 20 actions per conversation Flex Credits are cheaper and above it the flat conversation price is. Most support agents sit well under 20, which is worth knowing before someone signs up for the simpler-looking option.

A worked example, with every assumption on the table

Take a customer support agent handling 50,000 conversations a month, where a conversation takes six billable actions: authenticate the customer, look up an order, check a policy, generate an answer, update a record, send a summary. On Copilot Studio that maps to roughly one generative answer, four agent actions and one grounded retrieval, which is 2 + 20 + 10, so 32 credits per conversation.

The custom row is the one to distrust, and we are the wrong people to ask about it, because we sell custom builds. Its numbers are assumptions rather than published rates: three cents of model tokens per conversation, $1,200 of hosting, vector store, logging and monitoring, a $90,000 build amortised over twenty four months, and two engineer days a month of upkeep. Change the model or the context size and the token line moves several times over. It also excludes the risk that the build takes longer than planned, and the months before it does anything at all.

Where the number actually comes from

Look again at the Copilot Studio row. Ten of those 32 credits are tenant graph grounding, one design choice, costing $4,000 a month at this volume. Switch the questions that do not need tenant-wide retrieval to classic answers and the same agent drops to around 22 credits, or $8,800. Nobody changed vendor. Somebody changed the agent.

This is why the build versus buy framing misleads. The spread between a well designed and a careless agent on one platform is larger than the spread between platforms, and Microsoft makes the point themselves by publishing worked examples in the same shape: their support agent example computes as four classic answers plus two generative answers per run, priced per customer per day. The lever is the composition of a conversation, not the logo on the invoice.

Two operational details that belong in the same decision. Copilot Studio enforces prepaid capacity: at 125% of it, custom agents are disabled and users are told the agent is unavailable, so capacity planning is a live availability concern rather than a billing one. And the reasoning models everyone wants are billed on a second meter on top of the feature rate, so "let it think harder" is a pricing decision too.

Where each option actually wins

Run the same assumptions at 2,000 conversations a month and the picture inverts. Copilot Studio costs about $512, Agentforce on Flex Credits about $1,200, and the custom build still costs $6,610, because amortised build and infrastructure do not shrink with volume. At low volume the platforms are not slightly better, they are five to ten times better, and they work in weeks rather than months.

On these assumptions the crossover sits near 29,000 conversations a month, where the custom build's $6,550 of fixed monthly cost stops being worse than $0.256 per conversation. Below it, buy. Above it, the per-action meter starts to cost more than owning the thing. Your own crossover will differ, and the calculation is worth ten minutes with your real volume and your real action count.

Volume is not the only thing that flips it. Three others do, regardless of arithmetic:

• Write access to systems of record. If the agent must change data in something that is not the platform vendor's own product, you are building integration work either way, and the platform's advantage narrows to its interface.
• Logic the platform cannot express. Every one of these tools has a shape it expects. Fighting that shape costs more than writing the thing directly, and you find out late.
• Exit cost. Ask what you keep if you stop paying. Prompts and flows built inside a platform generally do not travel, which is a real cost that never appears in a comparison table.

What to do before anyone quotes you

• Write down one representative conversation as a list of actions. Not a user journey, the actual billable steps.
• Multiply by your monthly volume, then by each vendor's published rate. It is arithmetic you can do in a spreadsheet in an afternoon.
• Redo it with the expensive step removed, and see what your design is costing you.
• Ask what happens to that number if volume triples, and whether you hit a capacity wall before you hit a budget one.
• Only then ask whether to build.

If you want that worked through against your real numbers rather than ours, that is part of what a free AI audit covers, and we will tell you when the answer is to buy. Our engagement model explains how we scope a build when it is not. Related reading: what custom AI development costs and what to ask a vendor before signing.

Related reading

AI vendor security questions: what a good answer sounds like

The standard questionnaire will not tell you what you need. SOC 2 covers how a vendor runs their own company, not what the agent they build for you is able to do when it is talked into the wrong action. The questions that separate a serious builder are about permissions, credentials, and what capability they gave up to get there.

What you should own when the AI build ends

The code is the least valuable thing you funded. Name the prompts, the eval set, the embeddings and index configuration, the fine-tuned weights if any, and the infrastructure definition, in usable formats. The eval set matters most, because it is the artifact a successor cannot cheaply rebuild.

Can an AI project be fixed price? Only after discovery

Not the build, and yes the discovery. Fixed price requires knowing what correct looks like before work starts, and on an AI project that knowledge is the output of the first phase rather than an input to it. Fix the price of discovery, and let it produce the evidence a build number can stand on.