Home  /  Insights

Buying the tool before understanding the process

November 14, 2026 · Organisational Improvement

The US Department of Veterans Affairs began replacing its electronic health record in 2020, paused the rollout in 2023 after more than 800 major performance incidents, and restarted it in 2026 at four sites in Michigan. The original estimate was $10 billion; current estimates run from $16.1 billion to nearly $50 billion. What the VA changed before restarting is the most useful part of the story, and it was not the software.

The request arrives in a recognisable form. We have selected the tool. We need help with implementation.

Sometimes that is exactly right — the organisation knows its processes, has decided what it wants, and needs capable hands. More often the sentence is carrying something it should not. The tool has been selected to decide a question that nobody has yet answered: how this organisation is actually going to work.

That is not an implementation problem. It is a design problem that has been converted into a procurement, and the conversion is hard to reverse, because a purchase creates a deadline, a budget, a vendor with a delivery plan, and a political commitment that the choice was correct.

The case

In 2020 the VA began deploying a commercial electronic health record — the Cerner system, now Oracle Health — to replace VistA, the record system it had developed in-house over decades. The programme is one of the largest health IT undertakings anywhere, intended ultimately for about 170 sites.

The early deployments went badly, in ways that are unusually well documented because the VA has an active Office of Inspector General and a Congress that holds hearings.

At Mann-Grandstaff VA Medical Center in Spokane, Washington, the system routed more than 11,000 veterans’ clinical orders into what became known as the “unknown queue” — a holding area for orders the system could not match to a destination. The OIG found this directly harmed at least 149 veterans. The VA has since addressed it.

In Columbus, Ohio, in 2022, a system error meant a missed appointment was not routed to a queue prompting rescheduling. A patient died roughly seven weeks later from an accidental overdose. The OIG was careful about attribution: it found that staff failures compounded the software problem, including requests that were not evaluated and communications that were not sent, and cited a psychologist who did not thoroughly evaluate or address the patient’s depression.

In February 2024 a VA official told lawmakers that around 250,000 veterans were at risk of receiving contraindicated medications because of interoperability problems between the new and legacy systems. The OIG attributed this to an error in software coding, fixed within a week — but the patch did not resolve incorrect prescription data sent before it. Affected patients had not been notified, and the VA’s notices lacked outcome measures. The system also produced prescription backlogs, workarounds and pharmacy staff burnout.

In April 2023 the VA entered a reset, pausing new go-lives. The OIG reported more than 800 major performance incidents by September 2024, over half of them after the pause began. A GAO report in March 2025 found that only 13 per cent of VA staff using the system believed it made the VA as efficient as possible, and 58 per cent believed it increased patient safety risks.

The rollout resumed in 2026, going live at four Michigan systems — Ann Arbor, Battle Creek, Detroit and Saginaw — on the same day, with further deployments following in Ohio, Kentucky, Indiana, Alaska and Cleveland. Ten sites are live; full deployment means 170, with completion expected as soon as 2031. Secretary Doug Collins called the earlier rollout an “unmitigated disaster”. Deputy Secretary Paul Lawrence has said the system is “reliable”, with ten of the previous twelve months incident-free.

The part that matters

Now the useful question. What did the VA change during the three-year pause?

The headline change was standardisation: the VA limited local customisation, which Secretary Collins identified as a key problem, and standardised the deployment process itself. It replaced several competing governance councils with a single council reporting to top VA leaders. It expanded the programme office substantially, hiring towards 400 staff against an office of about 250.

Read that list again. Limit local variation. Standardise the deployment. Single governance line. Adequate staffing.

Not one of those is a software change. They are all answers to the question how will this organisation work — the question that a tool selection cannot answer and routinely appears to.

This is the whole lesson of tool-before-process, and it is why the pattern is so expensive. The organisation does not avoid the process design work. It defers it, into a phase where every design decision is now constrained by a configuration, a contract, a go-live date and a system already running in production at several sites. Design work costs perhaps five times more there than it would have cost up front, and that is before counting the harm done while the answer was missing.

Why the pattern is so common

It is not stupidity, and treating it as stupidity will not help you prevent it.

A tool is purchasable and a process is not. There is a budget line for software and a procurement process to spend it through. There is rarely a budget line for “six weeks finding out how we actually work”, and the person who proposes one is asking for money with no deliverable a finance committee recognises.

A tool decision feels like progress. Months of inconclusive process discussion end the moment someone signs a contract. There is a vendor, a plan, a timeline. The relief is real and it is why the decision is often made slightly early, by people who are tired of the discussion.

The vendor has a reference process and it is genuinely good. Modern enterprise products embed a sensible default way of working, often better than what the organisation does today. This is true and it is the trap, because the question is never whether the reference process is good in general — it is whether it fits this organisation’s constraints, which are frequently statutory, clinical, regulatory or safety-critical, and not visible in a demonstration.

Nobody owns the difference. The vendor owns the product. The integrator owns the configuration. Who owns the decision that a clinic will now book appointments differently? In most programmes, nobody — and that decision is the actual change.

The diagnostic

If you are being asked to implement a tool and you want to know, quickly, whether the process work has been done, ask these five questions. They take an afternoon.

Can somebody show me the current process, as performed, for the three highest-volume transactions? Not the documented process. The performed one. If no one can, the baseline does not exist, and without a baseline you cannot tell afterwards whether anything improved.

Where the tool’s standard way differs from ours, who decided which wins — and is that decision written down? This is the central question. In a healthy programme there is a register of these decisions with named owners. In an unhealthy one the answer is “we’ll handle that in configuration”, which means the decision will be made by whoever is in the room on the day, usually a configuration analyst with no authority to make it.

What are we deliberately not customising, and who said so? The VA’s own answer, arrived at three years late, was that local customisation had to be limited. Someone senior must own that boundary before build starts, because every individual customisation request will be reasonable and the aggregate will be unmanageable.

Which decisions can only be made by people who are not in this programme? Clinical, legal, regulatory, safety. If those people are consultees rather than decision-makers, their decisions will be made by default and discovered during testing.

What happens to work the tool cannot handle? Every system has an edge. The unknown queue is what an unowned edge looks like at scale. Name the exceptions, design where they go, and name the human who checks that destination daily. This is the question most often skipped and most often catastrophic, because an exception queue nobody owns is invisible until it is a serious incident.

Sequencing that costs less

Four to six weeks before the tool is configured, in roughly this order.

Map what happens now, for the highest-volume transactions, by observation rather than interview. Two or three weeks. Expect to find that the documented process and the real one diverge substantially; this is normal and is itself valuable information.

Separate the essential from the accidental. Which parts of the current way exist because of statute, regulation, clinical safety or a genuine customer need — and which exist because of a system being replaced, a reorganisation in 2014, or one person’s preference? This distinction is the entire value of the exercise. Essential constraints must shape the configuration. Accidental ones should be let go, and the tool’s standard way adopted.

Decide, and write it down. For each divergence: we adopt the standard way, or we customise, and here is why, and here is who said so. A short register. Twenty to forty entries is typical. This document is worth more than any other artefact the programme will produce.

Design the exceptions. Where does work go when it does not fit? Who looks at it, how often, and what is the escalation if the queue grows? Build the report that shows the queue depth before go-live, not after.

Done properly this costs four to six weeks and a handful of people. Set against a programme where estimates have moved from $10 billion to somewhere between $16.1 billion and nearly $50 billion, the arithmetic is not subtle — and it holds at every scale. I have run the equivalent exercise for a service management tool implementation in three weeks and it changed the configuration decisions substantially, which is the point.

The customisation argument, settled

Every tool implementation has the same fight, usually around month three, and it is worth knowing in advance how it resolves.

One camp says customise: our business is different, the standard product does not fit, forcing our people into somebody else’s workflow will destroy productivity. The other says configure only: customisation is what makes upgrades impossible and total cost explode, and every organisation thinks it is different.

Both are right about something and the argument is unwinnable as posed, because it is being conducted at the level of principle. It becomes tractable the moment you make it specific, and the divergence register is the device that does that.

For each divergence, ask one question: what breaks if we do it the standard way?

If the answer names a statute, a regulator, a clinical safety requirement or a contractual commitment, that is an essential constraint. Customise, and record why, so that in five years somebody can tell whether the reason still holds.

If the answer is that people will find it unfamiliar, or that it is more clicks, or that a particular manager prefers the current report format — that is an accidental constraint. Adopt the standard way and spend the money on change support instead, which is cheaper and does not compound at every upgrade.

If the answer is “I don’t know”, that is the most valuable result, because you have found a practice nobody can justify. These are common. In a typical mapping exercise a quarter of the steps turn out to have no defensible rationale, and removing them is pure gain regardless of which tool you buy.

The VA reached a version of this conclusion the expensive way. Limiting local customisation was named as one of the key changes before the restart — which is to say that after six sites, three years and several hundred performance incidents, the organisation concluded that much of its local variation had been accidental rather than essential. That finding was available in advance, for the price of a mapping exercise.

The Indian version of this

The pattern here usually arrives through one of two doors.

The first is the group mandate: a parent company or a global head office standardises on a platform and the Indian entity is told to adopt it. The process design was done elsewhere, for a different regulatory environment, and the local obligations — statutory record-keeping, local labour and tax requirements, RBI or SEBI expectations around audit trails and data — surface during user acceptance testing, when the budget is spent and the date is fixed. The defence is to get local regulatory constraints into the divergence register before configuration, treated as essential rather than negotiable.

The second is the licence deadline. A renewal, a discount expiring at quarter end, a bundled offer. A procurement decision taken on a commercial deadline forces a design decision onto the same deadline, and design work does not compress well. If the commercial offer is genuinely good, take it — and then protect the design weeks explicitly in the plan rather than letting the saved money be spent on a date.

What I would say to a sponsor

If you have already bought the tool, nothing above is wasted. The sequence still works; it is simply more expensive and more constrained than it would have been. Run it anyway, before configuration freezes. Four weeks spent now will be recovered several times over in the rework you do not do.

And if you are about to buy, one question. Ask the person asking for the signature:

When the tool’s standard way and our way disagree, who decides, and where will I be able to read that decision?

If there is a crisp answer with a name in it, the design work has been done. If the answer is that it will be worked out during implementation, you are not buying a tool. You are buying a deadline, and the design work is still ahead of you — at the worst possible price.


Sources

  • Federal News Network, VA EHR rollout resumes after three-year pause, April 2026 (2023 reset, Michigan go-lives, 10 of 170 sites, cost range $16.1bn–$50bn, standardisation and governance changes, Collins and Lawrence quotations, GAO March 2025 figures) — federalnewsnetwork.com
  • Nextgov/FCW, Issues with VA’s new EHR have contributed to patient harms and one death, watchdog says, 22 March 2024 (unknown queue, Columbus scheduling error, pharmacy interoperability and the 250,000 figure) — nextgov.com
  • Healthcare Dive, OIG details errors with VA, Oracle EHR deployment — healthcaredive.com
  • Military.com, VA halts future launches of its Oracle Cerner health record system, 21 April 2023 — military.com
  • Nextgov/FCW, VA watchdog lists EHR modernization as a major management challenge, January 2026 — nextgov.com
  • Figures and findings are as reported by the VA Office of Inspector General, the GAO and the sources listed, at the time of writing. The programme is ongoing and the position continues to change; verify current figures before citing them.

Leave a note