Enterprise AI in 2026: Access Is Solved, Depth Is Not

Deloitte's State of AI in the Enterprise 2026 report lands on a number that looks like victory and reads like a warning. Workforce access to sanctioned AI tools grew by roughly 50 percent in a year, moving from under 40 percent of workers to around 60 percent. Two-thirds of organisations, 66 percent, report gains in productivity and efficiency. On the spending side, Gartner forecasts AI agent software rising from 86.4 billion dollars in 2025 to 206.5 billion in 2026 and 376.3 billion in 2027, while projecting worldwide AI spending to grow 47 percent in 2026 alone. Then comes the counterweight: research from Writer finds 79 percent of organisations still face challenges translating adoption into value. Access is solved. Depth is not.

Access up, depth lagging

1. Access Is Not Usage

Deloitte's most useful distinction is between access and everyday usage. Access to sanctioned tools grew about 50 percent in a year, from under 40 percent to roughly 60 percent of workers, but the report explicitly notes that a smaller group uses those tools day to day. A licence is a permission, not a behaviour. The organisations capturing value are the ones that changed a workflow, not the ones that distributed seats.

# adoption_telemetry.json - one event per real use, not per licence
{
  "event": "ai_tool_invocation",
  "user_id": "u_48291",
  "tool": "coding_agent",
  "workflow": "ticket_triage",
  "outcome": "resolved",
  "human_review_minutes": 4,
  "tokens_in": 18400,
  "tokens_out": 2100,
  "cost_usd": 0.42,
  "ts": "2026-09-14T09:12:03Z"
}

2. The Spend Curve Is Steepening

Gartner's May 5, 2026 forecast puts AI agent software spending at 86.4 billion dollars in 2025, 206.5 billion in 2026, and 376.3 billion in 2027. On May 19, 2026 the firm projected worldwide AI spending to grow 47 percent in 2026. Whatever else you conclude, the budget is no longer experimental. When a category nearly triples in two years, the failure mode shifts from 'we tried it' to 'we cannot account for what we bought'.

# governance_manifest.yaml - depth requires an owner and a baseline
workflow: ticket_triage
owner: platform-team
baseline:
  cycle_time_minutes: 22
  cost_per_ticket_usd: 3.10
  reopen_rate: 0.14
controls:
  data_classification: internal
  pii_handling: redact_before_send
  human_in_the_loop: on_low_confidence
telemetry: s3://ai-telemetry/ticket_triage/

3. Why Pilots Stall

Writer's finding that 79 percent of organisations face challenges is not a model-quality problem. It is a process, governance, and measurement problem. Pilots stall when there is no baseline to compare against, when no one owns the workflow after rollout, and when the success metric is 'people liked it' rather than a business number. None of those failures require a better model to fix, which is exactly why they persist.

-- value_query.sql - did the investment actually pay?
SELECT
  date_trunc('month', ts)                        AS month,
  COUNT(*)                                       AS invocations,
  AVG(cycle_time_minutes)                        AS avg_cycle_min,
  SUM(cost_usd + human_review_minutes * 0.75)    AS total_cost_usd,
  SUM(CASE WHEN outcome = 'resolved' THEN 1 ELSE 0 END)::float
    / COUNT(*)                                   AS resolve_rate
FROM ai_telemetry.tool_events
GROUP BY 1
ORDER BY 1;

Platform architecture, not one-off pilots

4. Investment Moves to the Platform Layer

Analyst work from Ecosystm predicts enterprises will pull investment away from embedding AI into individual products and toward platform architecture, product management, AI governance, explainability, telemetry, and operational resilience. That is the depth gap expressed as a budget line. The organisations that struggled with one-off pilots are the same ones now buying registries, gateways, and observability, because a thousand unmeasured integrations is not a platform.

# rollout.yaml - expansion is gated on measured depth
stage_1:
  teams: [platform]
  exit_criteria:
    weekly_active_ratio: ">= 0.55"
    resolve_rate_delta: ">= +0.10"
    cost_per_outcome_delta: "<= 0"
stage_2:
  teams: [platform, support, data]
  requires: stage_1_passed_for_30_days
kill_switch: true          # every rollout keeps an off button

5. A 90-Day Depth Plan

Pick one workflow, not a portfolio. Write down the baseline number before you touch anything: cycle time, cost per unit, error rate. Assign one accountable owner. Instrument the workflow end to end so the metric moves for a reason you can name. Code sample 1 is the telemetry schema that makes that possible; code sample 3 is the query that answers, monthly, whether the investment paid. Expand only after one workflow moves.

# roi.py - the arithmetic everyone needs and nobody enjoys
def cost_per_outcome(invocations, tokens_usd, licence_usd, review_hours):
    outcomes = sum(1 for i in invocations if i["outcome"] == "resolved")
    if outcomes == 0:
        return None
    total = tokens_usd + licence_usd + review_hours * 75.0  # loaded hourly rate
    return total / outcomes

# Report this against the pre-AI baseline, or it means nothing.
# A number without a baseline is an opinion with a decimal point.

6. What to Measure

Three numbers survive contact with a finance review. Adoption depth, meaning how many of the licensed seats actually use the tool in a normal week. Value per workflow, meaning the measured delta against baseline. And cost per outcome, meaning tokens, licences, and human review time divided by the result produced. Code sample 4 is the rollout config that gates expansion on those numbers, and code sample 5 is the arithmetic nobody likes but everybody needs.

7. The Compliance Layer Nobody Budgets For

Depth almost always runs into the same wall: governance that was defined for a pilot does not survive production. Data classification, PII handling, retention, and human-in-the-loop review have to be specified per workflow, not asserted in a policy PDF. Gartner's own framing of agent spending alongside layoff and returns pressure is a reminder that the board will ask what the money bought. The workflows that scale cleanly are the ones where the compliance answer existed before the rollout, written down, owned by a named person, and testable. If your answer to 'where does this data go' is a shrug, the pilot is already over.

8. Signals You Are in the Depth Gap

Six signals. Your licence count grew and your weekly active ratio did not. Nobody can name the baseline number a rollout was supposed to beat. Success is reported as sentiment rather than as a metric. Every team has its own tools and none of them share telemetry. Cost per outcome is unknown, not merely high. And the last three pilots ended with a demo rather than a decision. Any two of those together mean the problem is not model capability, and no new model release will fix it. The fix is unglamorous: one workflow, one baseline, one owner, one number.

Measure value per workflow