Engageably

05 — Journal

← Back to Journal

AI translation is the discipline most engagements skip

May 20, 2026 · 8 min read · Studio · Benjamin Rodriguez

Executive TL;DR

  • Most AI engagements jump from model selection to deployment. They skip the work of turning outputs into the language and decision formats operators actually use.
  • Value leaks in the gap between what a model produces and what an executive can act on. The leak rarely shows up on the dashboard.
  • This is not a technical problem. It is interpretive, and it deserves a named discipline with a funded owner.
  • Firms that treat translation as a phase, not an afterthought, cut decision time and show return on AI spend their peers cannot.

The Macro Picture

AI adoption in large firms is now close to universal. Financial return is not. About 78% of organizations use AI in at least one business function, up from 72% in early 2024 (McKinsey, 2025). In the same study, more than 80% report no tangible EBIT impact from generative AI. Only about 6% qualify as high performers (McKinsey, 2025). Adoption is no longer the moat. Absorption is.

The failure mode is not model quality. MIT’s NANDA group studied enterprise generative AI programs. It found 95% of pilots deliver no measurable P&L impact, against an estimated $30 to $40 billion in spend (MIT Media Lab Project NANDA, 2025). The researchers call the cause a learning gap: systems that do not retain feedback, adapt to context, or fit into daily workflows (MIT Media Lab Project NANDA, 2025). A separate global survey found 60% of leaders say AI investments have produced little material value in revenue or cost (MIT Sloan Management Review, 2026).

This is a problem of practice, not tooling. The engagements that close the gap have a translation layer. The ones that do not, do not.

My Thesis

AI translation is the discipline of turning model outputs into decision-ready formats for a specific operator in a specific seat. It is not a deliverable at the end of an engagement. It is the engagement. MIT Sloan now calls this the last mile of AI, the point where a working model fails to become a used model in everyday decisions (MIT Sloan Management Review, 2026). We have worked this way for years. The field is catching up to the name.

The firms that staff, fund, and govern translation as a discrete capability will pull away. The model is not the asset. The decision the model changes is the asset.

The model producing a correct answer and the executive being able to act on it are two entirely different events. Only one of them creates value.

Default Behavior Versus Effective Practice

Default behavior ships outputs in the model’s native format. Ranked lists. Probability scores. Generated summaries. The executive is expected to interpret, route, and act. The handoff is treated as the finish line.

Effective practice runs the engagement backward. Map the decision context first. Who decides. On what cadence. With what adjacent data. Under what legal, brand, or financial constraints. Then engineer the output format from that decision architecture back to the model, not the other way around. A film director does not block a scene to show off the camera. The shot is designed to land one moment with one audience. The camera serves the moment. The model serves the decision.

The difference does not show up in model performance metrics. It shows up in whether the output changes a decision inside the window it was designed to influence.

P&L Impact in the Next 12 to 24 Months

Skipping translation does not show up as one line-item failure on a finance review. It shows up as quiet erosion across the operating P&L: slower cycles, repeat rework, stalled adoption, AI spend that cannot prove return. The pattern matches what MIT NANDA calls the learning gap, where systems that do not adapt to context never integrate into workflows (MIT Media Lab Project NANDA, 2025). The cost is real even when it is hard to attribute.

1. Decision delay cost. Output an executive cannot interpret on contact adds review cycles. Those cycles burn senior time and delay the action the model was supposed to speed up. Across a function, the compounding is material even when no single delay is.

2. Rework and re-engagement spend. Engagements that skip translation generate a second wave of spend when outputs come back for reinterpretation or reformatting. The original contract rarely scopes that second wave. It almost always lands on the buyer.

3. Adoption attrition. When executives cannot act on AI output without friction, use drops. The investment depreciates faster than the contract term. The impairment is real even when accounting treatment lags. MIT NANDA found only 5% of custom enterprise AI tools reach production (MIT Media Lab Project NANDA, 2025).

4. Competitive compression. Firms with a working translation layer make faster decisions from the same models their peers are buying. The gap is operational tempo. Tempo decides who responds to the market first.

Structural Risks and Governance Gaps

Translation Failure as Compliance Exposure

When the firm does not translate AI output into a decision-ready format, the executive never sees how confident the model is or where its limits are. They act on outputs they cannot accurately evaluate. The EU AI Act addresses this directly. Article 14 requires high-risk AI systems to let people oversee them effectively, understand their capabilities and limits, and avoid over-reliance. Article 13 requires technical measures that let deployers interpret outputs (EU Artificial Intelligence Act, 2024).

The financial exposure is not theoretical. Administrative fines for violations of transparency, documentation, and human-oversight obligations can reach €35 million or 7% of worldwide annual turnover (Legal Nodes, 2026). Translation is a compliance question with a price tag, not a style preference.

Accountability Gaps at the Handoff

If no role owns translation, two things happen. The executive who receives the output over-trusts it. Or a subordinate who cannot resolve the ambiguity quietly discards it. Neither outcome is auditable. The risk is not what the model produced. The risk is the missing human role that keeps the meaning intact between the model’s output and the decision.

Governance Without Translation Is Incomplete

Some AI governance frameworks cover data, bias, and model risk but ignore the translation layer. Those frameworks are structurally incomplete. They govern how output gets produced, not how it gets interpreted. Interpretation is where consequential decisions actually get made. A program can pass every model-risk review and still fail every decision it was built to support.

Operating Model and Talent Shifts

Firms are hiring at both ends. Model fluency on one side: prompt engineers, ML engineers, data scientists. Executive AI literacy on the other. The seat between them is largely unnamed and unfunded.

The translation function needs three things at once. Domain knowledge of how the receiving seat actually makes its decisions. A working understanding of how the model produced the output and where it is uncertain. The communication discipline to put both into a format that does not force the executive to hold all three at once. That is a senior practitioner. Not a soft skill bolted onto technical delivery, and not a junior role rebadged with an AI prefix.

Staff this as a core practice. Anything less leaves the operator doing the translation themselves, in real time, on consequential decisions.

The Playbook

Four moves close the gap in a single planning cycle. None require new technology. All require a named owner.

1. Audit the last three AI deliverables for decision readiness. Owner: CAIO Horizon: 30 days KPI: Share of outputs that required executive reinterpretation before a decision.

2. Name the translation function in every engagement scope.

3. Map decision architecture before model selection. Owner: CPO Horizon: Pre-engagement KPI: Decision delay, measured at output receipt.

4. Add a translation review to AI governance checkpoints. Owner: COO Horizon: Next governance cycle

Executive Next Step

Pick one active AI initiative this week. Ask who owns the translation from model output to executive decision. If the answer is no one, or the executive themselves, the program is exposed on return, on compliance, and on tempo. Name the owner before the next deployment ships.

Sources

← Back to Journal