Enterprise AI agents in 2026: what's shipping, what's a pitch deck

Every big software company now says it sells "AI agents". Salesforce, Microsoft, Amazon, Google, IBM, OpenAI — same word, different products, and almost none of it aimed at a five-person business. This post explains what the race actually consists of, using the figures the companies themselves have published, and separates what is genuinely in production from what is still a slide.

First, what an "agent" is supposed to mean

The word has been stretched to breaking point, so here is the useful version. A chatbot answers a question. An agent is given a goal, then takes actions on its own — looks something up, writes to a system, sends a message, books a thing — and reports back. The difference is whether the software can do anything or only say something.

That distinction matters commercially. A chatbot deflects a support ticket. An agent closes it. One of those saves you reading time; the other saves you a person's hour. Every announcement you are about to read is a bet that the second thing is now reliable enough to sell.

The numbers that are actually real

Most coverage of this race is vibes. There are a handful of published figures worth knowing, mostly because they tell you where the money is going rather than how good the technology is.

Salesforce: from five agents to thirteen

Salesforce publishes an "Agentic Enterprise Index" based on production activity across 400 businesses. Its finding: the average number of AI agents deployed per organisation went from 5 in early 2025 to 13 by April 2026. Time to create an agent fell by 53%. Employee sessions with agents tripled.

Read that carefully. "Deployed per organisation" is not "replaced a person". Thirteen agents across a company with thousands of staff could mean thirteen narrow jobs — an internal HR lookup, a lead-routing rule, a returns triage flow. The tripling of employee sessions is the more telling number: the main users of these agents are staff, not customers. This is internal tooling, not a robot workforce.

Salesforce's own product, Agentforce, went from roughly $500m in annualised recurring revenue in Q3 FY2026 to $1.2bn by Q1 FY2027 — two quarters, 205% year-over-year growth on their reported figures. That is a genuinely fast-growing product line. It is also a rounding error against Salesforce's total revenue, and it is sold almost entirely into companies that already pay Salesforce for something else.

IBM and OpenAI: consulting, with models attached

IBM and OpenAI announced a strategic partnership to embed OpenAI's models — GPT-5.6, Codex, ChatGPT Work — into IBM Consulting Advantage, with IBM standing up a dedicated OpenAI practice.

Translate that out of press-release language: IBM is training consultants to install OpenAI's models inside large organisations and bill for the work. That is a services business, not a product you can sign up for. If your company does not already have an IBM consulting relationship, this announcement changes nothing about your week.

Amazon: agents inside the advert

Amazon introduced "Alexa+ Agentic Ads" — a shopper asks a question, gets a personalised answer, and buys without leaving the ad unit. This one is the most concrete of the lot, because the business model is obvious. Amazon makes money from ads and from purchases. An agent that shortens the gap between question and checkout earns its keep immediately.

It is also a warning about where this is heading for small sellers. If more buying decisions happen inside an assistant's answer, the thing being optimised stops being your website and starts being whatever data the assistant trusts. Nobody can tell you yet what that means for a local business. Anyone who claims they can is selling something.

The rivals who agreed on a plug socket

The most interesting item is the least flashy. Anthropic, Google, OpenAI and Microsoft — four companies competing hard — jointly donated their competing agent-interoperability protocols to a shared, vendor-neutral foundation. They are betting that an interoperable ecosystem grows the overall market faster than any one of them winning a proprietary standard.

Think of it as agreeing on the shape of the plug before arguing about who makes the best kettle. It happened with USB, with HTTP, with shipping containers. It usually signals that the people involved believe the market is big enough that being first to a standard matters less than the standard existing at all.

It also quietly admits something: agents are only useful when they can reach your other systems. A clever agent with no access to your calendar, your CRM and your inbox is a very expensive autocomplete.

Shipping versus pitch deck

Here is the honest sort, based only on what has actually been announced.

Notice the pattern. The things with real numbers attached are sold to organisations that already spend six or seven figures a year with the vendor. The things that might reach a small business are standards and ad formats, and both arrive later, through someone else's product.

Why every one of them announced this at once

Three reasons, none of them mysterious.

1. Seat-based pricing is under pressure. Software has historically been sold per user per month. If software starts doing work instead of helping a person do work, the number of seats stops being the right meter. Every vendor is scrambling to invent pricing that survives that shift — per conversation, per outcome, per agent.
2. Agents justify the model spend. Training and running frontier models is enormously expensive. Chat subscriptions alone don't cover it. Agents that complete billable work inside enterprises are the business case being tested.
3. Nobody wants to be the one that didn't. Once two large vendors say "agentic", the other three have to, or analysts write that they're behind. A meaningful share of announcements in any technology cycle are defensive.

That third point is why a sceptical read is the right read. The announcements are real. The framing around them — that every business is about to run a digital workforce — is marketing doing its job.

What this race does not tell you

It tells you nothing about reliability in your situation. An agent deployed at a bank has a tightly scoped task, a QA team, a rollback plan and a compliance review. You have none of those, and you also have none of the overheads. Different problem entirely.

It tells you nothing about cost at your scale. Enterprise pricing is negotiated, bundled and invisible. The published figures are revenue totals, not price lists.

And it tells you nothing about whether your customers will tolerate it. That question is being settled in the wild, not in a vendor's index — and the answer is more nuanced than either side pretends. We wrote about the pushback in some customers really don't want to talk to an AI, and it remains the most important constraint on any of this for a business where people know your name.

Where the enterprise agent story genuinely isn't for you

Be direct about this. If you run a plumbing firm, a salon, a two-room clinic or a small agency, essentially none of what the big five announced in the past year is something you can act on.

• You don't have the systems to connect. Agent platforms earn their value by reading and writing across CRM, ERP, ticketing and data warehouses. If your CRM is a WhatsApp group and a notebook, there is nothing to orchestrate.
• You don't have the volume. Automating a process that happens 400 times a day pays back fast. Automating one that happens twice a week usually costs more to set up than it saves.
• You don't have the implementation budget. The licence is rarely the big number in enterprise software. The integration project is. That cost doesn't scale down with your headcount.
• You can't absorb the failure modes. A large company can tolerate an agent getting 2% of cases wrong because a team catches them. If an agent mishandles one of your eight enquiries this week, that's a real customer and a real review.

There's a flip side, though, and it's the genuinely useful takeaway: you don't need to wait for any of this. The parts of agent technology that work well today are narrow and well-defined — answering a phone, taking a message, filling in a form, drafting a reply for you to approve. Small tools that do one job already exist, cost tens of pounds a month rather than tens of thousands, and don't require a consulting practice to install. We looked at the current state of one such market in AI voice agent platforms in 2026.

How to read the next twelve months of announcements

You will see a lot more of these. A short filter that will save you time:

1. Is there a revenue figure or a usage figure? Salesforce published both. Most announcements publish neither. No number usually means no customers yet.
2. Is it a product, a partnership or a practice? A product you can buy. A partnership is two companies agreeing to try. A practice is consultants.
3. Who is the named customer? "A leading global bank" is not a reference. A named logo with a described outcome is.
4. Does it require something you already own? Most enterprise agent tooling is an upsell on an existing contract. If you're not already a customer, the price is different and usually much higher.
5. Is the word 'agent' doing real work in the sentence? Swap it for "feature". If the sentence still makes sense, it was a feature.

Apply that filter to the five items in this post and you get: one real product with real revenue (Agentforce), one real measurement of real deployments (the index), one consulting arrangement, one live ad format, and one standards agreement that will matter more in three years than it does today. That's a healthy, early, genuinely-happening market — and an entirely different market from the one a small business shops in.

The useful mental model is the one from the early web. Large companies spent millions on enterprise content management systems in the late nineties. Small businesses waited, and eventually got something simpler, cheaper and good enough from an entirely different set of suppliers. The same shape is playing out now, and the small-business end of it is already open. If you're curious what the simple end looks like in practice, what actually happens when an AI receptionist answers is a step-by-step version with none of the enterprise scaffolding.