The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale

Launching a system Launching a system proves its proves its proves its existence, but it doesn't existence, but it doesn't existence, but it doesn't prove that the people behind it prove that the people behind it will will use it. Or use it. Or use it. Or that she will ever that she will ever that she will ever make at least a dollar of make at least a dollar of make at least a dollar of profit from the profit from the profit from the deal model. I'm Michelle Grant, deal model. I'm Michelle Grant, deal model. I'm Michelle Grant, author of the Leading at author of the Leading at author of the Leading at Scale course, and the question facing Scale course, and the question facing every business right now is every business right now is this: the agent this: the agent this: the agent works, but does works, but does works, but does anyone trust him? I have anyone trust him? I have anyone trust him? I have spent over 20 spent over 20 spent over 20 years years years transforming transforming transforming businesses and over a businesses and over a businesses and over a decade decade decade leading leading leading integrations in integrations in integrations in mergers and mergers and mergers and acquisitions in media, acquisitions in media, acquisitions in media, technology, and technology, and technology, and consumer products. In consumer products. In consumer products. In these deals, trust these deals, trust these deals, trust wasn't something we wasn't something we wasn't something we just hoped for just hoped for just hoped for after launch. We after launch. We after launch. We built it before built it before built it before day one thanks to day one thanks to day one thanks to early early early change and risk management. change and risk management. So when the platform So when the platform went live, the teams went live, the teams went live, the teams on the ground really on the ground really on the ground really relied on it. AI relied on it. AI agents require agents require agents require exactly the same exactly the same exactly the same approach. An agent approach. An agent approach. An agent gains trust when he gains trust when he gains trust when he delivers the value delivers the value delivers the value planned by the business while planned by the business while staying strictly staying strictly staying strictly within the within the company's approved risk appetite, policies, and regulatory norms. In my company's approved risk appetite, policies, and regulatory norms. In my experience, leadership experience, leadership experience, leadership comes to this comes to this comes to this by answering three by answering three by answering three simple questions. What simple questions. What simple questions. What value does it create?

value does it create? value does it create? Who is responsible for Who is responsible for Who is responsible for it? And what evidence it? And what evidence it? And what evidence gives him permission gives him permission gives him permission to operate with greater to operate with greater to operate with greater autonomy? Once autonomy? Once autonomy? Once the agent exits the the agent exits the the agent exits the test environment test environment test environment into the main business into the main business process, the entire process, the entire process, the entire nature of the test nature of the test nature of the test changes. The pilot changes. The pilot changes. The pilot tests tests tests the functionality of the functionality of the functionality of the the the software, but software, but software, but the day-to-day work is where the day-to-day work is where the day-to-day work is where we verify whether we verify whether we verify whether it creates value it creates value it creates value within the controls within the controls within the controls set by set by set by management. management. Established AI risk frameworks give you strong AI risk frameworks give you strong principles. What they principles. What they principles. What they don't provide is how to don't provide is how to don't provide is how to apply these apply these apply these principles to each principles to each principles to each agent. For each agent. For each agent. For each agent, management agent, management agent, management must decide four must decide four must decide four things. What is he things. What is he things. What is he allowed to do? Who allowed to do? Who allowed to do? Who is its owner? What is its owner? What is its owner? What evidence does he record and evidence does he record and evidence does he record and what is the trigger for what is the trigger for what is the trigger for escalation. This work escalation. This work is not an engineering is not an engineering is not an engineering task. We believe task. We believe this is an executive this is an executive this is an executive operational mandate. And operational mandate. And operational mandate. And today I will show you today I will show you how to do it.

how to do it. Let's look at Let's look at this with a real this with a real this with a real example. Imagine a example. Imagine a $15 billion acquisition with $15 billion acquisition with the goal of generating $400 the goal of generating $400 the goal of generating $400 million in net million in net million in net synergies. The deal has already synergies. The deal has already synergies. The deal has already been closed, and been closed, and been closed, and integration efforts are now integration efforts are now integration efforts are now underway. The underway. The Integration Management Office, which we will call IMO today, will call IMO today, manages the manages the manages the synergy roadmap across 40 synergy roadmap across 40 synergy roadmap across 40 work areas. This work areas. This work areas. This includes everything from includes everything from includes everything from streamlining streamlining supplier contracts to supplier contracts to consolidating ERP consolidating ERP systems. And every systems. And every systems. And every initiative has a initiative has a initiative has a business owner business owner business owner who is responsible for its who is responsible for its who is responsible for its financial financial financial performance. performance. Every week, analysts at Every week, analysts at IMO will reconcile the numbers IMO will reconcile the numbers IMO will reconcile the numbers with the general with the general ledger. While collecting ledger. While collecting reports from all 40 reports from all 40 reports from all 40 managers, they managers, they managers, they also compare also compare also compare actual figures with actual figures with actual figures with planned figures, and then planned figures, and then planned figures, and then prepare a report for the prepare a report for the prepare a report for the steering committee.

steering committee. Which will include the Which will include the CFO, CFO, CFO, CRO, CRO, CRO, heads of business heads of business units, and units, and units, and head of integration. The head of integration. The engineering team engineering team creates an agent to creates an agent to creates an agent to automate this automate this automate this first draft. The first draft. The agent then reads the agent then reads the weekly updates, weekly updates, weekly updates, reconciles the register, reconciles the register, reconciles the register, calculates calculates calculates deviations from the plan, deviations from the plan, deviations from the plan, and flags and flags lagging initiatives. Everything lagging initiatives. Everything looks good, and looks good, and looks good, and they decide to they decide to they decide to run a run a run a four-week four-week four-week pilot project. pilot project. pilot project. The agent's drafts The agent's drafts The agent's drafts match the match the match the analysts' reports analysts' reports analysts' reports line for line. So, line for line. So, line for line. So, we can say that we can say that we can say that the technology works. the technology works. Now the question arises: Now the question arises: what value does the what value does the what value does the steering steering steering committee have to see to entrust an committee have to see to entrust an committee have to see to entrust an agent with the management of the agent with the management of the agent with the management of the business? First, the business? First, the business? First, the steering committee steering committee steering committee needs a written needs a written value proposition agreed upon value proposition agreed upon before the agent before the agent before the agent touches the working touches the working touches the working data. In our data. In our data. In our example, it example, it example, it comes down to three comes down to three comes down to three specific specific specific metrics. First, metrics. First, metrics. First, let's look at cycle time.

let's look at cycle time. If it currently takes five business takes five business days to prepare a weekly report, you'll want days to prepare a weekly report, you'll want days to prepare a weekly report, you'll want to set a to set a to set a target of, for example, target of, for example, target of, for example, two days. Your second two days. Your second two days. Your second metric will be metric will be analyst productivity. If analyst productivity. If four analysts four analysts four analysts spend half a spend half a spend half a week creating week creating week creating this report, the agent this report, the agent this report, the agent returns 80 hours per returns 80 hours per returns 80 hours per week. This time is week. This time is only considered valuable when the only considered valuable when the integration manager integration manager integration manager directs it to directs it to directs it to work streams that are work streams that are work streams that are lagging behind lagging behind lagging behind synergy goals and synergy goals and synergy goals and need help. need help. Furthermore, value Furthermore, value should only be considered should only be considered should only be considered after deducting after deducting after deducting the costs of computation, the costs of computation, the costs of computation, licensing, and licensing, and human verification of results. Your third metric is Your third metric is the most important—it’s the most important—it’s the most important—it’s detection time. This is the time detection time. This is the time detection time. This is the time from the first from the first from the first signs of a deviation in the signs of a deviation in the signs of a deviation in the data to the moment the data to the moment the steering committee sees it.

Reducing the report from Reducing the report from five to two days does not five to two days does not five to two days does not guarantee that the committee guarantee that the committee guarantee that the committee will see this will see this will see this deviation sooner. deviation sooner. The agent has to The agent has to capture it, and the capture it, and the value justification measures this value justification measures this directly. Each directly. Each directly. Each indicator requires a indicator requires a indicator requires a baseline before baseline before baseline before deployment. If deployment. If deployment. If the integration office (IMO) the integration office (IMO) the integration office (IMO) cannot name the cannot name the cannot name the source data source data source data today—don't today—don't today—don't deploy. If you deploy. If you deploy. If you can't measure where you can't measure where you can't measure where you started, you'll started, you'll started, you'll never prove never prove never prove that the agent made a that the agent made a that the agent made a difference. difference. Next, the steering Next, the steering committee needs to know committee needs to know committee needs to know who is who is who is responsible for responsible for responsible for the risks created by the the risks created by the the risks created by the agent. We agent. We agent. We will apply the will apply the will apply the three lines of three lines of three lines of defense model to the agent. At the first defense model to the agent. At the first defense model to the agent. At the first line of management, line of management, line of management, he is responsible for he is responsible for he is responsible for control. The control. The control. The integration manager is the integration manager is the integration manager is the responsible responsible responsible person. She person. She person. She is responsible for is responsible for is responsible for the results of the the results of the the results of the agent's work in the same way as she was agent's work in the same way as she was agent's work in the same way as she was responsible for the data responsible for the data responsible for the data collected manually. On the collected manually. On the collected manually. On the second line, second line, second line, risk risk risk and compliance departments and compliance departments and compliance departments oversee the oversee the oversee the agent and agent and agent and review review review his or her work methods. The depth of his or her work methods. The depth of his or her work methods. The depth of their vetting their vetting their vetting depends on the depends on the depends on the agent's risk level. An agent's risk level. An internal draft, internal draft, reviewed by a senior reviewed by a senior reviewed by a senior analyst, requires analyst, requires analyst, requires very different very different very different standards than an standards than an standards than an autonomous agent autonomous agent autonomous agent publishing figures publishing figures publishing figures directly to the directly to the directly to the board. At the third board. At the third board. At the third line, internal line, internal line, internal audit provides assurance by audit provides assurance by audit provides assurance by deciding, based on deciding, based on deciding, based on material risks, material risks, material risks, when and how an agent when and how an agent when and how an agent falls into the annual

falls into the annual falls into the annual audit plan. The agent audit plan. The agent audit plan. The agent also needs one also needs one also needs one employee employee employee assigned to him assigned to him assigned to him for daily for daily for daily supervision. In our supervision. In our supervision. In our example, this is a senior example, this is a senior example, this is a senior analyst. He analyst. He analyst. He reviews the draft reviews the draft reviews the draft before it before it before it goes to goes to goes to committee. Our role committee. Our role committee. Our role requires four requires four requires four specific specific specific capabilities. capabilities. First, she must First, she must understand the operational understand the operational understand the operational boundaries. She must boundaries. She must boundaries. She must interpret interpret interpret the results the results the results taking into account the taking into account the taking into account the context. It context. It context. It must have must have must have the power to the power to the power to repeal or repeal or repeal or amend them. And it amend them. And it amend them. And it needs a clear “ needs a clear “ kill switch” kill switch” kill switch” to safely to safely to safely stop the system stop the system stop the system if it behaves if it behaves if it behaves unexpectedly. unexpectedly. Accountability also Accountability also means defining means defining means defining where an where an where an agent's authority ends. agent's authority ends. I document this in a I document this in a one-page one-page one-page policy that I call the policy that I call the policy that I call the Human Human Human Performance Standard. For each Performance Standard. For each Performance Standard. For each branch of the process, branch of the process, branch of the process, it defines it defines it defines the decision and the role of the decision and the role of the decision and the role of the person the person the person authorized authorized authorized to make it.

to make it. to make it. Evidence needed and Evidence needed and Evidence needed and escalation path. These solutions escalation path. These solutions escalation path. These solutions exist at every exist at every exist at every level. The senior level. The senior level. The senior analyst decides whether the analyst decides whether the analyst decides whether the discrepancy is a discrepancy is a discrepancy is a timing difference timing difference timing difference or a real or a real or a real cash shortage cash shortage cash shortage . The workflow manager . The workflow manager decides how to address decides how to address this deficit. And the this deficit. And the this deficit. And the integration manager integration manager integration manager decides whether to officially decides whether to officially decides whether to officially flag the initiative flag the initiative flag the initiative as risky to the as risky to the as risky to the steering committee. steering committee. The agent prepares an analysis The agent prepares an analysis for each of these for each of these for each of these decisions, and that's where decisions, and that's where decisions, and that's where his role his role his role ends. The person ends. The person ends. The person who makes a decision who makes a decision who makes a decision is is is responsible for responsible for responsible for its consequences. So its consequences. So its consequences. So leaving these decisions leaving these decisions leaving these decisions to your team doesn't to your team doesn't to your team doesn't slow down the slow down the slow down the agent's work. It can agent's work. It can agent's work. It can work as long as each work as long as each work as long as each role retains its role retains its role retains its authority. The only authority. The only authority. The only reason to delay reason to delay reason to delay implementation is a implementation is a implementation is a solution without a solution without a solution without a designated designated designated owner. This means owner. This means owner. This means that unless management has that unless management has that unless management has agreed on who agreed on who agreed on who declares an initiative declares an initiative declares an initiative problematic, the problematic, the problematic, the agent's findings do not lead agent's findings do not lead agent's findings do not lead to action. The third question to action. The third question is where is where is where the engineers in the engineers in the engineers in this room come into play. Evidence for this room come into play. Evidence for this room come into play. Evidence for greater autonomy greater autonomy greater autonomy comes from how the comes from how the comes from how the agent is built. I am agent is built. I am agent is built. I am looking for three artifacts.

looking for three artifacts. The first is the The first is the audit log of audit log of audit log of decisions. For each decisions. For each decisions. For each workflow that workflow that workflow that the agent evaluates, it the agent evaluates, it the agent evaluates, it records the records the records the input data received, input data received, tool calls executed, tool calls executed, prompt and model versions, and prompt and model versions, and results generated along with human confirmation. Two engineering human confirmation. Two engineering choices make this choices make this choices make this magazine flawless. magazine flawless. First, your First, your basic basic basic math and reconciliation rules math and reconciliation rules math and reconciliation rules should be executed should be executed should be executed in deterministic, in deterministic, finance-approved finance-approved code, not code, not code, not generated as generated as generated as LLM colloquialisms. LLM colloquialisms. Second, treat the Second, treat the model's written resume as a useful model's written resume as a useful reference, not as reference, not as reference, not as proof. Your real proof. Your real proof. Your real proof is in proof is in proof is in the telemetry. What the telemetry. What the telemetry. What records were obtained? records were obtained? records were obtained? What deterministic What deterministic What deterministic tools tools tools were launched? And who were launched? And who were launched? And who signed the check? signed the check?

If the reported If the reported savings do not savings do not savings do not match the match the match the general ledger, general ledger, general ledger, the agent does not touch the figure. the agent does not touch the figure. It marks It marks the discrepancy and the discrepancy and the discrepancy and passes it on. passes it on. The second artifact is a model inventory record. It records the model inventory record. It records the responsible responsible responsible manager, manager, manager, data origin, data origin, data origin, dependencies between dependencies between dependencies between source systems, source systems, source systems, validation results validation results , escalation path, and , escalation path, and , escalation path, and instructions for the instructions for the instructions for the emergency stop. emergency stop. Most importantly, it Most importantly, it documents documents documents regression triggers—any regression triggers—any regression triggers—any updates to our updates to our updates to our base model or base model or base model or source API that source API that source API that require full require full require full revalidation revalidation revalidation before running in a before running in a before running in a production production production environment. Your environment. Your environment. Your third artifact is the third artifact is the autonomy scorecard. The steering committee steering committee uses it to uses it to uses it to grant autonomy grant autonomy grant autonomy in four stages. in four stages. The first stage is The first stage is shadow mode. shadow mode.

shadow mode. The analyst prepares the report The analyst prepares the report The analyst prepares the report as he does as he does as he does today. The agent today. The agent today. The agent works in parallel with works in parallel with works in parallel with it, and the team it, and the team it, and the team compares the results. compares the results. The second stage is a The second stage is a controlled controlled controlled draft version. draft version. The agent creates The agent creates a draft, and a senior a draft, and a senior a draft, and a senior analyst checks analyst checks analyst checks each line. The third each line. The third each line. The third stage is stage is stage is exception checking, where the analyst exception checking, where the analyst exception checking, where the analyst looks at flagged looks at flagged looks at flagged initiatives and a initiatives and a initiatives and a random sample of random sample of random sample of data. Next comes the data. Next comes the data. Next comes the fourth stage, where fourth stage, where fourth stage, where only only only exceptions remain. The analyst exceptions remain. The analyst exceptions remain. The analyst reviews only reviews only reviews only high- high- high- risk anomalies, risk anomalies, risk anomalies, supplemented by supplemented by supplemented by periodic periodic periodic independent sampling. independent sampling. However, pay However, pay close attention to the close attention to the close attention to the fourth stage, fourth stage, fourth stage, because when people because when people because when people only check for only check for only check for exceptions, your exceptions, your exceptions, your biggest biggest biggest vulnerability is what the vulnerability is what the vulnerability is what the agent failed to agent failed to agent failed to detect. Therefore, I advise detect. Therefore, I advise detect. Therefore, I advise tracking significant tracking significant tracking significant errors through errors through errors through independent sampling, independent sampling, independent sampling, controlling controlling controlling error severity, error severity, error severity, and closely monitoring the and closely monitoring the and closely monitoring the cancellation rate.

cancellation rate. Lowering Lowering cancellation rates doesn't cancellation rates doesn't cancellation rates doesn't automatically mean the automatically mean the automatically mean the agent is getting agent is getting agent is getting smarter. This smarter. This smarter. This may also mean that may also mean that may also mean that the depth of the check the depth of the check the depth of the check has decreased. So, an has decreased. So, an has decreased. So, an independent sample independent sample independent sample shows what exactly you are shows what exactly you are shows what exactly you are dealing with. The dealing with. The dealing with. The scorecard places scorecard places scorecard places these control these control these control metrics alongside metrics alongside metrics alongside value metrics value metrics : cycle time, : cycle time, analyst hours spent, and analyst hours spent, and discovery time. With discovery time. With discovery time. With this one- this one- this one- page overview, the steering page overview, the steering page overview, the steering committee can committee can committee can make clear, make clear, make clear, effective, and effective, and effective, and informed decisions informed decisions : promote : promote : promote an agent, leave things an agent, leave things an agent, leave things as they are, roll as they are, roll as they are, roll back, or back, or back, or return to the return to the return to the supervised supervised supervised review stage. So, let review stage. So, let review stage. So, let 's go back to the three 's go back to the three 's go back to the three questions. Place them questions. Place them questions. Place them on every agent on every agent on every agent you create, you create, you create, fund, or fund, or fund, or control. What control. What control. What value does it create? value does it create?

Who is responsible for it Who is responsible for it ? And ? And ? And what evidence what evidence what evidence allows him allows him allows him to act autonomously? In to act autonomously? In to act autonomously? In our our M& M& A example today, the agent didn't A example today, the agent didn't magically create $400 million in synergies magically create $400 million in synergies . . Procurement, Procurement, supply chain, and supply chain, and supply chain, and operations leaders have gained this operations leaders have gained this operations leaders have gained this value through value through value through painstaking painstaking painstaking integration work. But what the integration work. But what the integration work. But what the agent provides is agent provides is agent provides is obvious. obvious. Verified data Verified data weeks ahead of deadline, weeks ahead of deadline, weeks ahead of deadline, giving management time giving management time giving management time to intervene while to intervene while to intervene while synergies from the deal synergies from the deal synergies from the deal can still be realized. can still be realized. Creating a working Creating a working agent is just the agent is just the agent is just the first step. first step. The answers to these three The answers to these three questions are what questions are what questions are what earns trust for earns trust for