AI agents and the backlog: the case for adding capacity
Over the last few months I have reviewed a number of AI business cases, and it struck me how many of them start in the same place: how many roles can we remove? I understand why. Payroll is a number the CFO already trusts, and a saving against it is easy to build and easy to approve.
As a consultant I would boldly state that it is usually the weaker case. In recent operating-model work with a large enterprise, the leadership team made a deliberate choice to frame agents as a way to deliver more of the backlog with the expertise they already have. That choice changed the conversation about funding, about adoption and about what to measure. Here is why I think more organizations should make the same choice.
The opportunity cost sitting in the backlog
Anyone who has sat through an MBA economics module will remember opportunity cost: the value of the best alternative you give up. Most technology organizations carry a very large opportunity cost that never appears on a balance sheet. It is the backlog.
Ask any product or technology leader how much approved, valuable work is waiting for capacity. The answer is rarely “none”. Feature requests, regulatory changes, modernization, automation of manual processes, data quality fixes: the list is long, and most of it has a business owner who wants it done.
That backlog is where agents can produce the most value, because each item that moves from “waiting” to “delivered” carries its own benefit. A headcount saving is capped at the cost of the roles removed. The value of a larger delivered backlog is limited only by the quality of the backlog itself. So the question I ask leadership teams is simple: what is your backlog worth, and how much of it will you never get to?
The arithmetic, side by side
Take a simple team. These numbers are illustrative, chosen to make the comparison visible rather than to describe any client.
- Eight people deliver 40 backlog items a year.
- With agents working alongside them, each person delivers about a third more.
- A loaded role costs about $150,000 a year.
- An average backlog item is worth about $75,000 a year to the business once it is live.
There are two ways to use the extra capacity.
Option A: reduce the team. Six people with agents deliver the same 40 items. The saving is two roles, roughly $0.3M a year.
Option B: keep the team. Eight people with agents deliver about 53 items. The 13 extra items are worth roughly $1.0M a year.
Change the assumptions and the numbers move, but the shape usually holds. The saving in Option A is limited by the size of the team. The value in Option B grows with the quality of the backlog. If a team’s backlog is full of low-value work, that is a separate problem worth fixing first.
There are also costs that rarely appear in Option A’s spreadsheet. The people who leave take knowledge of the business with them. The people who stay learn that adopting agents shrinks their team, and adoption slows. Both reduce the capacity gain the saving was based on.
The capacity has to reach production
There is a catch, and it is the reason many AI pilots show strong developer numbers and weak business results. Mik Kersten makes the point well in Project to Product: value is only realized when work flows all the way through the value stream to the customer. Agents make the build step faster. They do not automatically speed up the review queue, the approval chain or the release window. In theory-of-constraints terms, optimizing one step is a local optimization, and the constraint simply moves elsewhere.
When I measured one large delivery estate, the automated steps were fast: builds started within seconds and the median approval took seconds. The waste sat in work that stalled with nothing to recover it: builds left running for days, approvals with no expiry, pull requests open for weeks. Adding agents to that system would have added more work to the same stalled queues.
So the capacity case comes with conditions:
- Put agents into review, test and release as well as build. An agent that prepares the evidence a reviewer needs shortens the review.
- Route changes by risk. Low-risk changes move through automated checks. People spend their review time on the changes that need judgment.
- Give every waiting step an expiry. Timeouts on builds, approvals and stale pull requests are configuration changes, and they are often the cheapest improvement available.
The roles stay and the work shifts
Framing agents as added capacity also makes the change easier to explain to the people doing the work. Nobody’s role disappears. What changes is how their time is spent.
The new skills sit on the top row: writing clear intent, checking agent output efficiently, and designing the checks that decide what is safe to release. Those are worth training for explicitly.
How to measure it
I get asked on many occasions which metrics prove that AI is working. My answer has not changed much since I first wrote about DevOps KPIs: define the business question first, and never look at a single measure in isolation. The business question here is “is the extra capacity reaching the business?”
Hours saved per developer does not answer it. Research from METR found experienced developers were about 19% slower with AI tools on their own codebases while believing they were about 20% faster, so self-reported savings are weak evidence on their own.
I prefer a small set of measures that are read together:
- Backlog items delivered per quarter, per team, compared with a baseline taken before agents arrive.
- Lead time from approved to live, which shows whether the queues are moving.
- Value delivered, using the benefit each item’s business owner signed up to.
- Cost per item delivered, including the agent and token spend.
The ratio that tells the story is value delivered relative to cost per item. If items delivered rise but lead time does not move, the work is piling up in a queue. If value rises while cost per item holds, the capacity case is working.
Take the baseline before the rollout. It cannot be reconstructed afterwards, and it is the evidence the next funding decision will ask for.
So, to conclude
Carefully think about what you want agents to do for your organization before you build the business case. If the goal is a payroll saving, the upside is capped and adoption will be harder. If the goal is to deliver more of a valuable backlog, the upside is larger, the people who know the business stay, and the measures are ones a CFO will recognize.
My suggestion is to start small. Pick one team with a long, valuable backlog. Measure its throughput and lead time for a month. Add agents to the build, test and review steps, put an expiry on its waiting steps, and measure again after a quarter. If delivered items and value go up whilst cost per item holds, you have a capacity case built on your own numbers.
This blog is my personal view and opinion, and is not shared by my employer.