Why do AI tools fail at law firms after they have already paid?

How we know this

Sidebar puts one question a week to legal management professionals at firms of 10 to 200 attorneys. Members are verified by title, employer, and firm size before they are admitted, and every reply is private. This page draws on every Sidebar cycle that has touched this question, and it is updated as new replies come in. We publish patterns across the group, never individual firms, and only once at least five members have replied to that question. Full methodology at gosidebar.ai/methodology.

Buying first, planning never

The most common failure is not a bad tool. It is a missing plan. Licenses get purchased under end-of-year budget pressure, with no governance policy, no training curriculum, and no rollout strategy, and the tool never gets activated. The purchase gets treated as the decision, when the decision that actually matters is how the firm will use the tool and what return it expects. Without that, even a capable tool reads as a failure within months.

A bad first impression does not reset

When a rollout stumbles, even briefly, attorney confidence drops and stays down. A pilot with early technical glitches can be fixed, but the damage to adoption often is not. Once attorneys decide a tool produces garbage, usually because no one taught them how to prompt it, the firm inherits a second and harder job: re-convincing skeptics to try it again. That recovery costs far more than getting the first rollout right would have. Vendors are not immune to this either. A flagship AI feature that visibly fails during a live sales demo, in front of an audience of prospective buyers, can leave a bad first impression that a stronger demo months later never fully shakes loose for the people who saw it stumble. The cost of a rocky launch is never contained to the firm that lived through it. It travels to everyone watching.

Purpose-built does not mean adoption-proof

The pattern holds even where firms did everything the playbook says to do. Among firms running NetDocuments, multiple members flag its bundled AI assistant specifically as immature and disappointing next to what attorneys already had for free in Microsoft Copilot. That is a useful correction to the instinct that a purpose-built, document-management-native tool must be the safer bet over a general assistant. It cuts against the headline finding above rather than confirming it, which is the honest state of this question: a general assistant can lose to a purpose-built tool on capability and still win on adoption, because the one attorneys already have open is the one they use. The tool being purpose-built for the domain does not exempt it from the same sequencing failures: a rollout without a training plan produces low adoption regardless of how well-matched the tool is to the task on paper. There is a specific trap hiding in the purpose-built assumption. A firm that reasons "this tool was built for exactly our workflow, so it will not need the same training investment as a general assistant" is skipping the one step that determines adoption. Domain fit changes how good the tool can be. It does not change whether anyone learns to use it.

Training on features misses the point

Firms that train users on how a platform works, rather than on the specific tasks attorneys find painful, see low adoption no matter how capable the tool is. The training that lands starts from two or three real jobs the attorney needs done and builds every session around those. The training that fails walks through menus and capabilities and leaves the attorney to connect it to their actual work alone. Most never do. The distinction is not about training quality in the usual sense. A well-produced session that walks through every menu and every capability can still fail completely, because a menu tour answers "what can this do" and never answers "what do I do with the thing sitting on my desk right now." Attorneys do not generalize from a feature list to their own workload on their own. That translation step has to be built into the session, or it does not happen.

Slowing down is the quiet winner

A small share of firms report no failed tools at all, and they credit one thing: they refused to move fast. Taking an extra month or two in evaluation led them to the right fit instead of the first fit, and they now carry no shelfware and no re-engagement problem. In a market that rewards the appearance of speed, the firms with the cleanest AI stack are the ones that were willing to look slow. That is a hard case to make internally, because slow evaluation looks identical to indecision from the outside, and a partner asking why the firm has not signed anything yet is not going to be satisfied by "we are being careful." The firms making this argument successfully are pairing the extra time with a visible plan and a deadline, so the delay reads as discipline rather than paralysis. An open-ended evaluation with no end date is indistinguishable from the paralysis this same group of members describes elsewhere as its own failure mode.

The failure rate our members describe is not unusually high

It is close to the industry norm. Harvard Business Review, drawing on MIT research into generative AI programs across industries, reported that roughly 95% of gen AI pilots fail to deliver a measurable return. That figure covers business broadly, not law specifically, but it reframes what our members describe: a firm with one failed AI rollout is not an outlier running behind its peers. It is close to the median. What separates the small share of firms getting a return, in law and elsewhere, is not a better model or a bigger budget. It is the unglamorous sequencing work our members keep naming: a plan before the purchase, training built around real tasks, and enough patience in evaluation to find the right fit before signing. None of those three things costs meaningful money. All three require someone with the standing to slow a purchase down, which is a scarcer resource inside most firms than the budget line for the license itself.

Source: Harvard Business Review, What Companies With Successful AI Pilots Do Differently

What to do with this

Before any new AI rollout, run one structured problem-framing session with the attorneys who will use the tool. Map two or three specific tasks they find painful today, then build every training touchpoint around those tasks rather than around the platform's feature list. That single step is what separates the firms in this group with no failed tools from the firms re-selling last year's purchase internally. It costs a meeting, not a budget line, and it has to happen before the license is signed, not after adoption has already stalled. If a rollout is already underway and stalling, run the same session retroactively. It is never too late to replace a feature tour with a task list, and doing it late still beats not doing it at all, even if it means re-earning trust the first launch already spent.

Frequently asked questions

Why do AI tools fail at law firms after the firm already paid for them?

Usually the tool is not the problem. The most common failure is a license bought under budget pressure with no governance policy, training plan, or rollout strategy behind it. Three to twelve months later the tool sits unused, and the firm blames the technology rather than the missing sequence around it, then quietly repeats the exact same pattern with the very next purchase.

How common is it for AI pilots to fail, in law and elsewhere?

Very common. Harvard Business Review, citing MIT research, reports that roughly 95% of generative AI pilots across industries fail to deliver a measurable return. A law firm with a failed AI rollout is closer to the median outcome than to an unusual or embarrassing outlier.

Does a purpose-built AI tool adopt better than a general assistant like Copilot?

Not automatically. Among firms running NetDocuments, many flag its bundled AI assistant as underwhelming next to Microsoft Copilot, which attorneys often already had for free. Being purpose-built for a domain does not exempt a tool from the same training and sequencing failures that sink general-purpose rollouts.

What is the single best step to prevent a law firm AI rollout from failing?

Run one structured problem-framing session with the attorneys who will use the tool before it is purchased, mapping two or three specific tasks they find painful today. Firms that build training around those real tasks, rather than the platform's feature list, report far higher and more durable adoption.