AI Agents Now Come With Consultants Attached. That Tells You Where the Hard Work Is.
IBM and Microsoft, and separately Cognition and McKinsey, are now selling AI agents bundled with consulting. That is a signal that the hard part of enterprise AI is the operating model, and engineering leaders should treat it that way.
AI Agents Now Come With Consultants Attached. That Tells You Where the Hard Work Is.
Two announcements landed within a week of each other, and I think they say more about enterprise AI than any benchmark released this year.
On November 10, IBM and Microsoft announced Imagine Enterprise, a platform to build, run and govern AI agents that starts with more than 6,000 prebuilt agents and assets, with financial services among its first industry frameworks. Six days later, The Next Web reported that Cognition and McKinsey formed an alliance to sell Devin, Cognition's AI software engineer, to large companies, with McKinsey choosing projects, setting baselines and redesigning teams around it.
Look past the logos and the pattern is clear: the agent is being sold together with the people who change how your organization works. I don't read that as a sales gimmick. I read it as an admission from the vendors themselves that the model is no longer the bottleneck. The operating model is.
The packaging is the message
Imagine Enterprise is not something you download. According to IBM's announcement, clients get it by engaging IBM Consulting, by co-innovating with IBM on larger implementations, or by self-managing, and forward-deployed engineers from both companies are offered on qualifying engagements. MarketScale's analysis put it well, describing the purchase as "closer to a delivery engagement than a software download." It also noted that the release counts agents and assets together without a split, and states no price.
The Cognition and McKinsey deal follows the same logic from the engineering side. Per The Next Web, McKinsey cited its own survey of 334 product and engineering leaders, which found top-quartile organizations reporting that AI more than doubled productivity, while others saw productivity fall. That spread is the whole story. Same class of tools, very different outcomes, and the difference is how the work was reorganized around them.
What 20 years of shipping software tells me about "6,000 agents"
At Priceline, I spent about twelve years on high-traffic travel systems. The lesson I carried out of that job is that the expensive part of any capability is never the first version. It is the integration, the edge cases, the on-call rotation, and the fifth change request after go-live.
A catalog of prebuilt agents is useful the way a catalog of open-source libraries is useful. It shortens the start. It does not tell you how an agent behaves when your customer data has twenty years of inconsistent history, when two systems disagree about the status of a request, or when the agent is confidently wrong at 2 a.m.
On my current team, we are embedding AI agents into a banking service ticket portal that handles more than 1,000 tickets a day, starting with triage, routing and follow-ups. We built it on Spring Boot, Spring AI and Azure OpenAI, with human approval on higher-risk actions. The model calls were the quickest part to get working. Most of our effort went to things no catalog could ship for us: defining which actions count as higher risk, deciding who approves them, making every agent decision traceable, and agreeing with operations teams on what "correctly routed" even means.
Buy, build, or partner: the MBA lens
My MBA coursework kept pushing one question I now ask about every AI investment: what are you actually paying for, and who holds the risk when it goes wrong?
With consulting-led agent platforms, you are paying for three things at once: software, accelerators, and change management. That bundle can be the right call. If your organization has never run an agent in production, borrowing experienced people may be cheaper than learning every lesson yourself. But I would go in with clear eyes on a few points:
- Price the outcome, not the engagement. If there is no published price, build your own cost-per-outcome model before signing, such as cost per resolved ticket or per completed code change, and compare it to your baseline.
- Own the knowledge transfer. Forward-deployed engineers are valuable only if your own engineers leave the engagement able to run, debug and change the agents without them.
- Watch the dependency surface. Imagine Enterprise leans on Microsoft's stack for identity, governance and security, including Entra, Purview and Defender. That may suit a Microsoft shop well. It is still an architectural commitment you should make deliberately, not by default.
- Keep accountability inside the bank. In regulated work, a partner can design controls, but the institution answers for the outcome. Human-approval rules and audit trails should be defined by your risk and engineering leaders, not inherited from a template.
What this means for software engineering teams
The Devin alliance matters for engineering leaders in a different way. The Next Web reported that Cognition says it works with six of the eight global systemically important banks. AI coding agents are already inside large financial institutions, and the question is shifting from "should we allow this?" to "how do we organize around it?"
My team uses GitLab Duo for AI-assisted development, and the biggest gains have come from process changes, not from the tool's raw output. A few things I would recommend to any engineering manager:
- Baseline before you roll out. Measure cycle time, review time and defect escape rate now, so you can tell later whether the tool helped or just moved work around.
- Treat code review as the new bottleneck. When an agent produces more code, reviewers absorb the load. Plan reviewer capacity and set clear rules on what an agent may change without senior sign-off.
- Redesign roles, not just workflows. Engineers who spend more time specifying, reviewing and testing need coaching for that work. That is a leadership job, and it should not be outsourced.
- Start where the blast radius is small. Test generation, documentation and well-bounded refactors are better first targets than core transaction logic.
The takeaway
Vendors bundling agents with consultants is an honest signal: the technology is good enough, and the hard part is everything around it. Identity, approvals, audit, cost tracking, team design and change management decide whether an agent becomes a durable capability or an expensive pilot.
Engineering leaders can buy help with that work. We cannot buy our way out of understanding it. Whatever partner you choose, make sure that when the engagement ends, your team knows how the agents work, what they cost, and when a human must step in.
Sources
- IBM and Microsoft Announce a New AI Agent Platform Designed for Industry-Specific Transformation — IBM Newsroom
- IBM and Microsoft announce a new AI agent platform designed for industry-specific transformation — Microsoft Source
- Cognition and McKinsey team up to sell Devin, its AI coder, to big firms — The Next Web
- IBM brings 6,000+ agents and assets to Imagine Enterprise — MarketScale
Enjoyed it? Tap to react. You can clap more than once.
New articles by email
Occasional notes on AI, engineering leadership and building teams. No spam, and you can ask me to remove you any time.