Black office chairs with chrome arms in a row at a conference table.

OpenAI’s consulting deals show the hard part of AI agents is permissions

OpenAI Frontier Alliances hand AI agent rollouts to big consultancies. Harper Singh sets out a permissions-first rule any mid-size firm can run itself.

A 60-person distributor wants an AI agent to sort supplier emails. The operations lead gives it a mailbox of its own, forwarded copies only, no send rights, and reads its drafts every morning for two weeks. Nothing gets deleted, because nothing can be. That’s our illustration, and it’s the whole lesson of the OpenAI Frontier Alliances in one office: permissions decide how an agent rollout goes, and you can set them yourself before you pay anyone.

On February 23 OpenAI announced multi-year alliances with BCG, McKinsey, Accenture and Capgemini. Its own page says what limits enterprise value is how agents are built and run, with model smarts second. That’s a vendor describing its own market in a launch that sells deployment services, so read it with that in mind. OpenAI is also reshuffling how it spends on safety, and the OpenAI training pause matters less than the monitoring budget.

Key points

  • The alliances pair OpenAI with four big firms building practice groups and certified teams, working beside OpenAI’s forward deployed engineers.
  • Reported agent failures come from what the account was allowed to do. A prompt asking an agent to behave can be lost. An account limit can’t.
  • Start with one workflow, one owner and an agent account that does only what you’d let a new temp do on day one.
  • Paid help earns its fee when the agent must work across your ERP, CRM and finance tools at once and a mistake costs real money.

What OpenAI is selling with the Frontier Alliances

Frontier is OpenAI’s platform for building and managing AI coworkers. It was open to a limited set of customers, with wider access promised over the following months, and OpenAI published no pricing. Each alliance partner is building a practice group around it.

Advertisement

OpenAI’s COO, Brad Lightcap, said at the India AI Impact Summit that the company has not yet really seen enterprise AI penetrate business processes, according to TechCrunch’s report. He said OpenAI would measure Frontier on business outcomes instead of seat licences. We read the alliances as an admission that connecting agents to real systems is slow work. Some platforms already refuse outright, and Amazon’s block on AI shopping agents shows what a refused permission looks like.

The inbox that shows why access matters

Summer Yue, a Meta AI security researcher, said her OpenClaw agent started deleting her real inbox and ignored stop commands from her phone. She believes the larger inbox triggered compaction, where an agent summarizes its own history, and that it dropped her instruction to stop. TechCrunch reported her account and could not verify it. It’s one self-reported incident, and the compaction idea is her hypothesis.

Her safeguard was a sentence in a prompt. A sentence can be dropped or ignored. If the agent’s mailbox can’t delete, it can’t delete, however confused the agent gets. Think of a contractor: you can ask them to stay away from the electrical panel, or you can keep the key.

And yes, the framing suits OpenAI. If the constraint is deployment, buyers need deployment help, and OpenAI now has four partners selling it. The inbox story suggests the framing is still partly right. We made the same point about personal agents in our Moltbot security checklist, where exposed gateways came from broad default permissions.

Your first rollout, written on one page

  1. Pick one workflow and name one owner. Skip the committee.
  2. For each system the agent touches, write one line: read, draft or act. Most first projects need only read and draft.
  3. Give the agent its own account. It never borrows your login. For email, it writes to a drafts folder and can’t send or delete.
  4. Put every irreversible action (send, delete, pay, publish) behind a human click outside the agent.
  5. Log what it does, and know how you’d cut its access in under a minute.
  6. Review the log after two weeks, then widen one permission at a time.

Choosing the workflow? Start where your staff already use AI, which our Gallup adoption column shows leaders tend to misjudge, and price it per accepted result as in our Sonnet 4.6 pricing column. The first project should feel dull. Dull is good.

When to call in a consultancy

When the agent has to span several core systems and a mistake costs real money, the four firms are built for that, and a company of a few hundred staff can reach that point. A single-workflow pilot at a 50-person firm usually isn’t that, and your own operations lead can run it.

The sceptic says a 200-person firm with no IT lead will get the permissions wrong, and a consultant is cheaper than one agent deleting the wrong folder. Reasonable. A narrow first project is still the cheaper test of whether you need help. If you can’t scope one workflow and one account in a week, hiring help then is a better-informed purchase. Before you sign anyone, choose an AI consultant by the last project they actually shipped, and ask what you’ll own at the end: the configuration, the permission map and the logs.

We can’t compare prices, because OpenAI published none, and we’ve seen no results from any alliance engagement. We haven’t used Frontier. The same caution applies to cheap models, which our column on AI model distillation for buyers covers. If firms start reporting agents working safely under broad access, we’ll drop the narrow-rights rule. We haven’t seen that yet.

Frequently asked questions

What are the OpenAI Frontier Alliances?

They are multi-year partnerships between OpenAI and BCG, McKinsey, Accenture and Capgemini to help large companies deploy AI agents on OpenAI’s Frontier platform. OpenAI announced them on February 23, 2026.

Does a small business need a consultancy to deploy AI agents?

Usually, no. A single workflow with one owner and a restricted account is a reasonable first step. Outside help makes more sense when agents must touch several core systems.

How do I stop an AI agent from deleting things?

Don’t rely on a prompt. Give the agent a separate account that cannot delete, and require a human click for anything irreversible.

We have not used OpenAI Frontier. Archive entry dated February 25, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.

Total
0
Shares
Prev
Anthropic’s distillation case is less about China than about the cheap model you’re about to buy
A rack of electronic server equipment in a dark room.

Anthropic’s distillation case is less about China than about the cheap model you’re about to buy

AI model distillation is back in the news after Anthropic's claims against three

Next
Nano Banana 2 is cheap per image, but the price you should track is per usable image
A black DSLR camera resting beside a camera lens.

Nano Banana 2 is cheap per image, but the price you should track is per usable image

Nano Banana 2 pricing starts near 7 cents an image

You May Also Like