"Beware of little expenses; a small leak will sink a great ship."

Robots

People

Love

Reading time: 3 min.

In Deloitte's new State of AI in the Enterprise survey of 3,500 senior leaders, 71% of companies track their total AI spend, but only 44% track the unit cost: the price of a single call, which shows what one task or workflow costs to run.

With agents, that gap gets expensive, because long reasoning chains, retries and agents calling other agents can multiply the compute behind a single business task, the report notes, and the AI can quietly end up costing more than the manual work it was meant to save.

Pick one agent workflow and trace what each finished task costs, step by step. Fixed steps, like matching an invoice to its purchase order, can often run for far less on the automation you already own, which is one reason Forrester expects enterprises to move 20% of agentic workflows back to plain rules in 2027. Then give each workload a named owner for its cost, as Deloitte recommends, and agree the number with Finance.

How human-centric AI delivers business results

Your weekly briefing on tools that create competitive leverage

Duvo maps how a process such as supplier bonus reconciliation works today, then carries out the version your team approves inside the systems you already have, sending exceptions to the person who owns them. It raised a $32M Series A two weeks ago, and the AI underneath can be swapped without rebuilding the process. At retailer Notino, quarterly bonus reconciliation went from 70 days to 15. Best for COOs who want a whole process automated while its owner keeps sign-off on changes.

Kestra is open-source software for fixed, repeatable workflows, written as code your IT team can review. An AI agent becomes one step inside the flow, with guardrails on what goes in and what comes out. Its 2.0 release governs agents and plain workflows side by side. Built for Chief AI Officers whose agents must run inside processes IT controls.

Online auction platform HiBid used to spend three weeks on its month-end close. After automating order-to-cash with Hyphenate, it takes one day. Ramnandan Krishnamurthy co-founded Hyphenate, which raised a $45M Series A this week, to run finance agents on top of a company's existing ERP and banking systems. The agents learn from past spreadsheets, emails and closes, and each resolved exception becomes a precedent for the next close, so the team's past decisions carry forward.

When teams decide which workflow goes to an agent next, many reach for a scorecard with five to ten criteria, each rated 1 to 10. Use four criteria, rated 1 to 4. Four covers what matters without the over-analysis, and the short scale reminds everyone these are rough opinions rather than hard data. For agent ideas, I'd put confidence on the list: how sure you are that every step holds up once it runs daily.

The LEAP Guide walks through a full rating session, and the templates are free on the site.

One book, handpicked from my conversations with friends, industry leaders, and tech innovators:

Donald Sull and Kathleen Eisenhardt looked at how people handle complex problems, from the Federal Reserve to Tina Fey's 30 Rock. In Simple Rules, they argue that a handful of rules of thumb, aimed at the bottleneck that matters most, often works better than long, detailed procedures. It makes a good companion for the week you decide which workflow steps run on rules and which need a person.

Lighthouse takes us many hours to create, and we don’t run ads or charge anything. We are dedicated to making it the best AI transformation newsletter possible.

We just ask for one thing: tell one friend about Lighthouse today!