> ## Content Index > Fetch the complete content index at: https://www.therecursive.com/llms.txt > Use this file to discover other available public pages before exploring further. # Enterprise AI Doesn’t Need Better Models But Better Workflows - URL: https://www.therecursive.com/enterprise-ai-doesnt-need-better-models-it-needs-better-workflows/ - Published: 2026-09-22T08:44:50.000Z - Updated: 2026-09-22T08:47:40.000Z - Description: Think your business is saving millions on AI? Think again. Comparing cheap token bills to human salaries is a trap quietly blowing up enterprise budgets. How to redesign workflows to measure what actually matters? - Author: Serhiy Tokarev - Tags: Guest Article, AI, B2B, corporate Most companies calculate AI ROI incorrectly. They compare the monthly token bill with an employee's salary and report "90% savings". This is a trap. The only metric that matters is the **cost of an accepted outcome**: a result that meets the company's requirements for quality, speed and acceptable risk. It includes inference, integrations, retries caused by errors, monitoring, human review, rework and the expected cost of risk. Companies that optimise token costs instead of using this formula can end up spending more on AI than they previously spent on people — without seeing any productivity gains. ## Where AI reduces the cost of an outcome The cost of an accepted outcome is lowest where two conditions overlap: **complex, unstructured inputs and an output that a human can verify within seconds**. Attempts to automate entire functions (“replace the analysts”, “automate the legal department”) almost always fail: the scope is too broad, while validating the output ends up costing as much as doing the work manually. We saw this in the initial screening of startups. Previously, prioritising a long list of potential investment opportunities was manual: an analyst opened each company and checked it against our investment criteria one by one. Now our internal assistant, built on Claude, runs the full list against the criteria that matter most at pre-seed and seed — team excellence and market opportunity — and **returns a ranked list with a short rationale for each company**. We then do deep research from the top of that list. It is the same funnel, but with a better order of operations: the analyst now focuses on the top 10 instead of reviewing 100 startups, and productivity at the initial screening stage has increased roughly tenfold. The same principle works in our **legal processes**. Previously, every NDA was handled by a lawyer from scratch: reviewing the provisions one by one, assessing risks and checking them against our established position, then making changes manually. **We changed the logic of the process itself**. We prepared a detailed handbook for the agent outlining the company's position on key NDA provisions — which terms are acceptable, which require changes, and which should be escalated to a lawyer. **The agent conducts the initial review**, prepares a short summary and proposes a redline, while the lawyer reviews and refines the prepared markup. This reduced NDA processing time from 1.5 hours to 15–20 minutes. ### It didn't all go smoothly The agent's first deployment also showed the limits of the approach. Initially, it applied the playbook too literally: if the wording of a contractual provision differed from the example in the playbook, it could flag it as problematic and replace it entirely with the standard language; even when human review showed that the substance of the provision already met the acceptable standard, just in different words. The same happened with edits: where a targeted change was enough, the agent would sometimes rewrite the entire provision. We fixed this by expanding the example base, adding more acceptable variations of the same language. The agent began to focus on the substance of the provision and learned to distinguish between cases where a targeted edit was enough and those requiring a full redraft. ## Human-in-the-loop as a risk-management tool When I analyse a company (whether it's a pre-seed startup in France or a potential Series A investment in the US) AI becomes my first analyst: reconstructing the competitive landscape, reviewing financing history, and looking for red flags. **Instead of spending thirty minutes collecting basic information**, I spend that time assessing the resilience of the business model. But I never ask AI whether I should invest. The responsibility and the signature remain mine. > This is the principle I consider critical for enterprise AI: understanding and action should be separated. AI can interpret complex information, identify patterns and prepare recommendations. But clear rules should determine what the system can do itself, where approval is required and how the decision is recorded. **Human review is not a separate philosophical position on AI safety**. It is a line item in the cost of an accepted outcome: the expected cost of an error is the same kind of component as inference or integrations. Removing a human from that line does not make the outcome cheaper, it simply hides its true cost until an error surfaces. That is why enterprise AI autonomy should be an outcome, not a starting condition. AI should move from reading information to acting independently within clearly defined limits, and each additional level of autonomy should earn trust only once it is backed by measurable reliability and audit trails. Until then, human-in-the-loop is part of the working system. ## The model is 10% of the product. The rest is integration Access to frontier models such as GPT-6 Astra or Claude Fable is not a competitive advantage on its own. The model is one component of the system: it can read a contract or process an invoice, but it cannot create a business outcome on its own. > **Without integration into internal systems, the cost of using it rises** — not in the subscription bill, but in the hours people spend stitching the process together manually. **LP communication is a good example**. We have more than 60 companies in our portfolio, and each portfolio update has its own set of materials: emails, presentations, updated financial models, reporting when available, plus updates on the fund itself. Previously, work with prospective LPs was fragmented: the team communicated separately with hundreds of contacts, and every new conversation meant assembling all this content again from scratch. Now we run a unified LP platform built with Claude Code. Fund materials, portfolio updates and reporting sit in one place, so a new LP conversation starts from a current, consistent base instead of a rebuild. Less time goes into assembling materials, more into the actual conversation; **bringing the data together in one context has removed the hidden time costs of reconciling versions**. ## Processes win, not models Our cases demonstrate one thing: the cost of an outcome falls not when a newer model appears, but when the process around it is redesigned. As an investor, I wouldn't consider access to a frontier AI model a long-term competitive advantage. Models quickly depreciate and become cheaper. The real edge comes from **proprietary domain data, deep integrations, evaluation systems, auditability and a product's position** inside the customer's daily workflow. Across our portfolio companies, AI agents don't replace teams — they extend them. Their value comes from specialised systems being orchestrated around a business process. The question today isn't whether AI agents will replace employees. The question is whether you have moved from optimising token costs to optimising the cost of an accepted business outcome; and **whether you have built a process in which AI can actually make that outcome cheaper**.