The AI Agents Already Inside Your Software: What Embedded Agents in ERP and CRM Mean for Buyers

Microsoft's new Microsoft 365 pricing took effect this week. In its December announcement, Microsoft put E3 up from 36 to 39 dollars per user per month and E5 from 57 to 60, and the stated rationale was that the suite now carries more AI and security capability than it did when the old prices were set. Nobody at most companies decided to buy that capability. It arrived with the license, and now it's priced into the license.

I've sat in a lot of renewal conversations this year, and the same realization keeps landing late: the AI agents people are evaluating from startups already exist, in some form, inside the ERP and CRM systems they own. If you buy software for a living, or you're the person who signs off on it, this changes three practical things. What you might be double-paying for. How new capability enters your environment. And what's negotiable at renewal. I'll take those one at a time, then give you an audit you can run in an afternoon.

How fast this is moving

Gartner expects 40 percent of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. Even if reality lands at half that, the direction is set, and the same Gartner research projects agentic AI could account for roughly 30 percent of enterprise application software revenue by 2035, more than 450 billion dollars. Vendors have every commercial reason to move fast here, because agents are the growth story their boards have already priced in.

A quick vocabulary note, since "agent" gets used for everything now. The agents shipping inside business software are mostly task-specific. One drafts the follow-up email from the call recording. Another triages the inbound ticket and proposes a resolution. In accounts payable, one matches the invoice to the purchase order and flags the exception. Salesforce ships these as Agentforce inside Sales and Service Cloud. Microsoft ships Copilot agents across Microsoft 365 and Dynamics. SAP's Joule has grown from a chat assistant into a family of dozens of task agents across finance, supply chain, and HR. The mid-market suites are running the same play at smaller scale, with NetSuite, HubSpot, Zoho, and the major helpdesk platforms all pushing agent features into regular releases. None of it waits for you to run a procurement process.

You may be paying for the same thing twice

Over the past two years most companies accumulated a layer of standalone AI tools. A meeting notes tool for the leadership team. An AI email assistant for the sales reps. A support chatbot from a startup. An invoice-processing tool for accounts payable. Each purchase made sense on the day it was made, and each one now has a sibling being built into a suite you already pay for.

Here's a hypothetical that will feel familiar. Say you run a 200-person distributor. You pay for a standalone note-taker at 20 dollars a seat for 30 managers, an AI sales-email tool at 50 dollars a seat for 12 reps, and a website support bot at about 1,000 dollars a month. That's roughly 2,200 dollars a month in standalone AI spend. Meanwhile your CRM tier now includes agents that draft follow-ups from call recordings, and your helpdesk vendor shipped a deflection agent in its spring release. Some portion of that 2,200 is buying capability you already own.

To be fair to the standalone tools, the embedded agent is often the weaker product. Plenty of point tools are still better at their one job than the suite's version, sometimes dramatically better. So the real question is a comparison across four axes: output quality on your actual work, fit with the workflow where the work happens, where the data already lives, and contract timing. When the suite agent is 90 percent as good and the data is already in the suite, consolidation usually wins. When the point tool is clearly better and the team lives in it all day, keep it and look for savings elsewhere. The failure mode I keep seeing is that nobody runs the comparison at all, so both line items renew on autopilot.

New capability now arrives through vendor updates

Traditional software procurement has a gate. A new tool means a new vendor, which means a security review, a data processing agreement, maybe a pilot, and a named person accountable for the decision. Embedded agents walk around that gate. They arrive in a quarterly release, sometimes enabled by default, sometimes one admin toggle away, and no purchase order ever triggers the review.

This matters because agents read and act on data by design. An agent that summarizes support cases reads your support cases. An agent that drafts customer emails reads the thread history. An agent that updates records can write to them. Every question your security review would normally ask, about data access, about where inference runs, about whether your data trains someone else's model, about what the agent may do without a human approving it, still applies. Those questions just never get asked, because nothing was bought.

Gartner has a related prediction that I read as a warning about exactly this. It expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Embedded agents make the risk-control part harder, because deployment can now happen without anyone deciding to deploy. Three cheap governance moves close most of the gap. Give each core system a named owner who reads the release notes monthly and flags agent features. Extend your AI acceptable-use policy to cover vendor-embedded agents, since most policies I've read only contemplate tools employees sign up for on their own. And treat enabling an embedded agent as a change that goes through the same lightweight review you'd give a small new tool purchase.

The agent line item is a SKU, and SKUs are negotiable

Vendors haven't settled on how to charge for agents, and unsettled pricing is negotiable pricing. Salesforce launched Agentforce at 2 dollars per conversation, then added a consumption model in May 2025 in which agent actions draw down prepaid Flex Credits, working out to roughly 10 cents per action sold in 500-dollar packs, alongside per-user bundles at 125 dollars a month for sales and service. SAP includes its base Joule assistant in cloud contracts but meters Joule Agents and premium AI scenarios through prepaid AI Units, a separate line on the order form. Microsoft went the other direction for part of its stack and folded capability into the base suite price, which is how you get this month's increase.

Consumption pricing deserves your closest attention, because an "action" is not a unit you can forecast from experience. Say your service team handles 4,000 cases a month and an agent touches each case three times. That's 12,000 actions, about 1,200 dollars a month at a dime each, which is manageable. But agents chain actions together. A single case might trigger a classification step, a knowledge lookup, a draft, a revision, and a record update, and now your multiplier is five or eight instead of three, and the invoice surprises you. Before signing anything metered, ask the vendor for a sample consumption report from a customer of your size, and negotiate a cap that alerts you rather than billing overage silently.

Other levers work right now. Pilot credits are widely available because vendors want referenceable agent customers and will fund proof periods. Rate protection on the meter keeps the per-credit price fixed for the term. A written definition of what counts as a billable action prevents the multiplier problem from being reclassified as your misunderstanding. And timing the agent conversation to your renewal, when your leverage peaks, beats negotiating it mid-term. If your account rep tells you agent pricing is fixed, that hasn't matched anything I've seen this year.

An audit you can run in an afternoon

You can capture most of the value of this post with one exercise. Block an afternoon and involve whoever owns IT spend plus one person who knows the systems well.

1. List your core systems. ERP or accounting, CRM, helpdesk, HRIS and payroll, plus the two or three platforms your operation actually runs on. For each, write down which agent features are included in your current tier today, which cost extra, and what the next two releases add. Release notes and public roadmap pages get you most of the way. For the rest, email your account manager and ask in writing, including what will be enabled by default.
2. List your standalone AI spend. Pull twelve months from the expense system and the corporate cards, because a chunk of this was bought by individual teams and never touched an IT budget. Subscriptions, per-seat tools, usage-billed APIs. Put a monthly number on each.
3. Map the overlaps. Match the second list against the first and sort into three buckets. Cancel or consolidate, where the suite version is good enough and the data already lives there. Keep and defend, where the point tool is clearly better, and write down why so the decision survives the next budget review. And included but unused, agent capability you're already paying for in a suite tier and using at zero, which is either free value waiting or an argument for dropping a tier.

While you're in the admin consoles, check which agent features are already switched on. In my experience there's usually at least one that nobody remembers enabling.

Questions to put to your vendors before renewal

• Which agent capabilities are included in my current tier, and which are a separate SKU? Get the answer in writing, because the marketing pages blur this line.
• What is the metering unit, what does a customer my size typically consume in a month, and will you share a sample consumption report?
• What data do the agents read and write, where does inference run, and is any of my data used to train models outside my environment?
• Can I set a hard consumption cap, and what exactly happens when I hit it?
• Which agent features in your next two releases will be enabled by default, and what admin controls exist to turn them off?
• If I consolidate a standalone tool into your agent, what commercial support will you offer for the migration?

What I'd do before the fourth quarter

Renewals concentrate in the fourth quarter and January, so the audit is worth running now, while there's still time to act on what it finds. Bring the overlap list to every renewal conversation. Ask the six questions above and file the answers with the contract. And add embedded agents to your change process, so the next capability that arrives by vendor update gets reviewed by someone with the authority to say not yet. At FirmAdapt we run a version of this exercise with clients regularly, and the overlap list has never once come back empty. I'd be surprised if yours does.