The third option is where almost every firm should be, and it is the one nobody puts on the slide.
Start a conversation with the AI Adoption Concierge, already scoped to build, buy, or configure. Pick a starting point, or describe your situation directly.
Firms tend to frame this as a binary and then choose badly. Buying is presented as unambitious, building as serious — and the evidence does not support that ranking. Research across enterprise AI deployments found internal builds succeeding at roughly a third the rate of purchases from specialist vendors, with failures clustering around integration, data quality and ownership rather than model capability. The option that gets left off the slide is configuration: shaping a vendor product or an open framework with prose, prompts, playbooks, templates and connectors rather than code. That is where firms without engineering capacity have produced genuinely differentiated results, and it is what the very large firm-built platforms are mostly doing at scale anyway — the largest disclosed programme in the market reportedly involved several hundred lawyers contributing requirements alongside its technologists, which says the scarce input is domain specification, not engineering.
Most firms need the middle two and reach for the outer ones.
Right where the capability is undifferentiated or the failure mode is malpractice-grade — research, e-signature, docketing, records retrieval, time capture.
The sweet spot. Shaping a product or open framework with firm-specific prompts, playbooks and templates. No code, real differentiation.
Joining systems you already pay for. The highest-leverage new category, and largely configuration rather than construction.
Justified only when all of: genuinely differentiating, no vendor serves it, permanent engineering capacity, an evaluation set exists, and a partner sponsor who will still care in eighteen months.
Every retrieval project dies here first. Scanned archives, inconsistent metadata, duplicates and no status field beat model choice every time.
If you cannot measure whether output is right, you have not deployed a tool — you have deployed an unaudited associate.
How to decide, and what to do first.
Mostly whether the firm ends up with a capability or a stranded asset.
Model versions are retired on published schedules, and 2026 saw a large family of widely-used models scheduled for retirement together. A firm with a validated prompt library or an integration pinned to a specific model has a migration on a deadline it did not set. Ask every AI vendor which model identifiers they call and when those retire — it is a fair question and a revealing one.
Writing down how your practice works, in prose, in a form a model reads before doing anything else — the firm's house style, its escalation thresholds, the clauses it will and will not accept, the questions it always asks at intake. Then assembling prompts and templates around that as reviewed, versioned firm work product rather than as individual habits. Several vendor platforms and at least one significant open framework are explicitly built around this pattern, and the artefacts are plain text. It is unglamorous, it needs an editorial owner rather than an engineer, and it is the most under-used lever in the profession.
Almost certainly not, and the market has moved against it — one major provider has been winding down self-serve fine-tuning entirely. The reasons are practical: fine-tuning teaches style and format rather than facts, so it does not stop a model inventing specifics; meaningful training needs thousands of labelled input-output pairs, which a firm does not have even if it has thousands of documents; and training on client material is precisely the scenario ethics guidance flags as requiring informed client consent. A well-engineered prompt with a handful of house-style examples plus retrieval over your own precedents gets most of the benefit with none of the consent problem.
You cannot, unless you built the evaluation set before you started — and this is the step that separates firms that get value from firms that get enthusiasm. Take thirty to fifty past matters where the correct output already exists in the file, have a senior lawyer confirm what makes an answer correct, and freeze it. That becomes the regression suite: run it before any model migration, after any prompt change, and periodically against tools that update silently. Without it, the firm is relying on impressions, and impressions in this field have been shown to diverge sharply from measured performance.
Not a spectacular failure — a quiet one. The pattern is a capable person builds something genuinely useful, it becomes load-bearing, nobody documents it, the underlying model is deprecated or the person leaves, and the firm discovers months later that it has been producing worse output or nothing at all. Every safeguard against that is administrative rather than technical: a named owner and a named successor, artefacts in a firm repository rather than a personal account, a scheduled regression run, and a calendared note of the retirement date of whatever model it depends on.
Describe what you are trying to achieve. The Institute will help you work out whether building is the answer.