Everyone Is Buying AI Agents. Should You?

Publish date

April 16, 2026

Publish date

April 16, 2026

You already know the question is coming.

Probably this quarter. Maybe next week. Someone with more authority than patience is going to ask what your AI agent strategy is, and “we are evaluating options” is not going to land the way it landed last year.

So you start evaluating. Four demos in three weeks. Three of them are good. One of them is so good it feels like a trap.

It probably is.

Half the Purchases Are Right. What about the Other Half?

Half the AI agent purchases happening this quarter are the right call. The other half are going to become somebody’s worst quarter of 2027. The hard part is that the two halves look identical in the demo.

Here is the split we see every week inside client conversations.

The right call is buying an off the shelf agent when the work is generic, the data is not regulated, and the vendor’s incentives line up with yours.

The wrong call is buying the same kind of agent when the work touches regulated data, business logic that lives in three people’s heads, or a workflow that cannot fail without a regulator getting involved. Same demo. Same pitch. Same contract shape. Completely different outcome.

If buying an off-the-shelf agent is the default, it must be the financially safe bet, right?

It isn’t. When you actually map the Total Cost of Ownership (TCO) of activating AI add-ons across a standard 1,500-employee tech stack, buying becomes a compounding financial trap.

Adding AI to your existing stack isn’t an upgrade. It’s a 60% markup to rent your own business logic.

By buying the AI modules pushed by your CRM, Workspace, HR, and Support vendors, a standard mid-market company absorbs a $1.85 million annual markup. You are paying a seven-figure penalty to rent fragmented logic that you do not even own.

What about the Cost?

JPMorgan built its own AI platform called LLM Suite and rolled it out to 230,000 employees. The bank banned ChatGPT from internal use before it built its replacement. More than 450 production use cases today, targeting 1,000 by end of year.

Here is what that play cost. A $19.8 billion technology budget in 2026. Roughly 2,500 AI specialists on payroll, up from 1,500 in 2022. Three years from the ChatGPT ban to the current rollout. Most financial services firms do not have $19.8 billion. Most do not have 2,500 specialists to hire. Most do not have three years.

Goldman Sachs embedded engineers from Anthropic inside the firm in February to build internal agents for accounting and client onboarding. Not a vendor contract. Engineers inside the building. That is a bespoke arrangement Goldman paid for at Goldman scale. It is not a product. It is not repeatable. A regional bank, an insurance carrier, or a midmarket asset manager cannot sign up for it.

Mayo Clinic cut readmissions by 40 percent using AI powered monitoring built on its own data. Cleveland Clinic runs a virtual triage system at 94 percent accuracy in its emergency departments, built for Cleveland Clinic, by Cleveland Clinic. Both institutions have internal data science teams measured in the hundreds. Most hospital systems have internal data science teams measured in single digits.

None of these institutions is behind the buy side curve. They are running a different curve, and the curve compounds in a way the buy side curve does not. The 230,000 employees using LLM Suite are training the next version of LLM Suite. The agents your competitor just bought have no equivalent feedback loop. The buyer pays the vendor to learn. The builder learns for itself.

The play is correct. The resourcing is not repeatable.

Most of the market cannot wait three years, hire 2,500 specialists, or embed vendor engineers inside the building. The version of this play that works at the scale you actually operate at is the one we build with clients every week. Same sovereign posture. Same feedback loop. A fraction of the cost and the timeline. That is what Mustang was built for.

Three Things the Builders Refuse to Outsource

The data itself. Not because of caution. Because of statute. HIPAA, GLBA, SOX, FedRAMP, OCC examination findings. These do not bend for vendor convenience. When a regulator asks where your customer data lives, “inside the vendor’s tenant” is not the answer that ends the conversation. It is the answer that starts a new one.
The exceptions. Off the shelf agents perform beautifully on workflows they were trained for and break the moment they touch a workflow that lives in somebody’s head. The credit officer who remembers why this borrower’s covenant package never got refiled in 2019. The compliance reviewer who flags certain claim types on Thursdays because the audit cycle runs Friday. An off the shelf agent cannot be told this. It was not built for it.
The roadmap. Every off the shelf agent is built for the median customer. When you need a feature changed because of an examination finding, you are in a queue behind two hundred customers who do not have that problem. When the vendor pivots, you pivot with them. When the vendor raises prices, you pay them.

None of these are arguments against the buy side. They are a description of who those products were built for. The category is winning because most enterprise workflows do not have these constraints.

Yours might.

Four Questions. Ask Them Before You Sign.

Before you sign, run the workflow you are about to automate through four questions. Two or more answers on the build side means buying is the wrong default for that specific workflow. Not for every workflow. For that one.

Does your data move under statute?

If a regulator can issue a subpoena, examination order, or privacy enforcement action against this workflow, the vendor’s tenant is the wrong place for it.

Are your edge cases written down?

If your hardest exceptions live in three people’s heads and cannot be captured in a runbook, you are not buying an agent. You are training one. That is build territory.

Does the vendor’s pricing reward the outcome you actually care about?

Outcome billing works when the outcome is clean. Resolved ticket, completed booking. It breaks when the outcome is regulatory. Clean audit, no fine, no incident report.

Can you walk away in 18 months?

If the year of prompts, workflows, orchestration, and learned exceptions you accumulate on top of the vendor does not leave with you, you are not buying software. You are signing a lease with no exit clause.

Most companies sit on a mix. Some workflows pass all four questions and the answer is buy. Some fail two or more and the answer is build. The mistake is not picking a side. The mistake is treating the whole portfolio as one decision.

Sort the Workflow. Do Not Pick a Side.

The companies getting this right in 2026 are not buying everything and they are not building everything. They are sorting.

Off the shelf agents for the public facing consumer support that does not touch protected data. Enterprise search for general knowledge work. Legal research for non privileged work. And a sovereign agent layer for the workflows where the data, the exceptions, and the regulator all live in the same room.

The sort is the work. The question is not whether to buy AI agents. The question is which workflows belong on which side of the line.

You already know the question is coming. When it lands, the answer is not yes or no. The answer is which workflows, and why, and what runs where.

If that is the conversation you need to be ready for, it starts with Mustang.