The H2 Operations Reset: How to Decide What to Automate for the Rest of the Year

Why mid-year is the right time to reset your operations

The turn from H1 to H2 is the one moment in the year when you have enough data to be honest and enough runway to act on it. You are six months into whatever plan you set in January, close enough to remember what actually ate the days and far enough from year-end that changes you make now compound for two more quarters. January resolutions are guesses. A mid-year reset is a review of evidence.

Most operations reviews go wrong because they start with ambition instead of accounting. The founder asks what to build or what to launch, when the more useful question is where did the hours go and which of those hours were spent on work that required no real judgment. The busywork that quietly reset your effective capacity in H1 is almost never on the roadmap. It hides in the gaps between the roadmap items: the follow-ups that did not go out, the report someone rebuilds every Friday, the reconciliation that slips to 9pm at month-end.

This piece is a practical playbook for that review. The goal is not to automate everything, and it is not to chase the most painful problem first. It is to look at your last six months clearly, sort the work into what a person must own and what an AI workforce can absorb, and set up the rest of the year so your team spends its time on the decisions that actually need them.

Step one: audit where the last six months actually went

Before you decide what to automate, you have to see the work, and most teams have never written it down. Spend an hour reconstructing a normal week from the systems you already have. Your inbox shows the messages that took a first reply and the threads that went cold. Your calendar shows the standing meetings and the recurring admin blocks. Your CRM shows the leads that got one touch and no follow-up. Your project tool shows the status updates someone assembles by hand. You are not doing time-and-motion accounting; you are building a list of the tasks that repeat.

For each recurring task, note three things in plain language: how often it happens, roughly how long it takes each time, and how much judgment it really requires. Be ruthless on that last one. A lot of work that feels demanding is actually just tedious and frequent, which is precisely the profile you want. Triaging inbound support, drafting first-pass replies, chasing an unpaid invoice, summarizing a call into a CRM note, pulling the weekly numbers into one place. None of it is glamorous and all of it recurs, which is exactly where hours leak.

The output of this step is a single ranked list: the repetitive work of your business, sorted by how many hours it consumes across a quarter. That list, not a wishlist of features, is the input to every decision that follows. If you skip this and jump straight to tools, you will do what most teams do, which is point AI at the loudest problem instead of the most automatable one, and then quietly give up when it does not stick.

Step two: sort the work — what an AI workforce can take, and what judgment keeps human

Take your ranked list and sort every item on two questions. First, how much judgment does it require: could a competent new hire do it from a one-page SOP, or does it need context, taste, and accountability you cannot write down? Second, how reversible is a mistake: if the output is wrong, do you catch and undo it cheaply before it reaches a customer or the books, or is it expensive and embarrassing to walk back? Those two axes do more to tell you what is safe to hand off than any feature comparison.

The automate-first work is high-volume, low-judgment, and reversible. It is the front half of tasks, not the whole task: reading and classifying, pulling context a human would otherwise dig for, and writing a first draft. An AI workforce is genuinely good at this now, and it is good at doing it on every item rather than just the easy ones. This is where an AI agent that works inside your own tools earns its keep, because the work already lives in your inbox, CRM, docs, and project tools, not in a new place you have to check.

The judgment stays human. Pricing exceptions, hiring and firing, legal commitments, the decisions that define a key relationship, anything expensive or one-way to reverse. AI can and should draft, research, and prepare these, because the legwork is still legwork. But the call belongs to a person, and a confident-sounding draft is not a substitute for a real decision. Kirality keeps a hard line here on purpose: money, communications, and legal always keep a human in the loop, no matter how much you trust the rest of the system. That line is not a limitation you graduate past. It is the point.

Step three: start human-in-the-loop and prove it before you trust it

Whatever you decide to hand off, start it in propose-and-approve mode. The AI does the reading, scoring, and drafting, and then a person reviews and approves each action with a click before anything happens. Nothing fires on the AI's own authority. This is not a temporary training-wheels phase you are impatient to leave; it is how you find out whether the agent actually understands your business before it is allowed to act on that understanding.

The reason to start here is that a bad draft is a non-event and a bad autonomous action is an incident. When the AI drafts a support reply that misreads the policy, the reviewer fixes or kills it in seconds and the customer never sees it. When an unsupervised system sends that same reply, you own it. Human-in-the-loop lets you get most of the speed of automation without inheriting its worst failure mode, and it pays the cost of being wrong in review time rather than in customer trust.

Prove it with numbers you decide up front. The one that matters most is your approval rate: the share of drafts you send with little or no editing. A rising rate means the agent understands that slice of your work and you can safely widen the lane. A flat one tells you exactly where it still needs help. Run the AI in parallel with your existing process for a week or two, watch what you approve untouched versus rewrite, and let the data, not your enthusiasm, decide when a task is ready to graduate. Every edit and rejection is also a signal that makes the next draft better, so you are improving the system by doing your normal job.

Step four: graduate the safe repetitive work to standing mandates

Once a task has earned trust — the approval rate holds, the failure modes are known and cheap — you can graduate it from click-by-click approval to a standing mandate. A mandate is a rule you set that lets routine work run within limits you define, so you are not personally approving the forty-first identical low-risk action. You are not handing over judgment; you are pre-approving a narrow, well-understood pattern and keeping the boundary tight.

The craft is in the limits. A good mandate is specific: which exact task, up to what threshold, within what scope, with what fallback when the agent hits its own uncertainty. Auto-categorize routine transactions under a dollar limit but route anything above it or anything unusual to a person. Send the standard appointment reminder on the schedule, but escalate a reschedule that touches a refund. The mandate should describe the boring middle of a task and hand the edges back to a human by default. Reversibility, not judgment, is what decides whether a step is safe to let run — so anything hard to undo stays behind a click.

Some things never graduate, and that is deliberate. Money going out, communications sent under your name to people who matter, and anything legal keep a human in the loop permanently, regardless of how well the surrounding automation performs. This is also a security property, not just a taste one: keeping a person on consequential steps is what stops a prompt-injection attack hidden in an incoming email from quietly tricking an agent into sending money or exfiltrating data. And because every proposal, approval, and mandated action is written to a tamper-evident, append-only audit trail, you can always answer who or what decided a thing, and when — which is what makes widening autonomy a decision you can defend rather than a leap of faith.

A function-by-function starting map for H2

Support triage is the cleanest first lane for most teams. Let the AI read, classify, and route every incoming ticket, pull the account context a human would otherwise spend minutes digging for, and draft an on-brand first reply grounded in your real docs and past resolved tickets. Keep sends human at first. Once your approval rate on a given ticket type is consistently high, a standing mandate can let the agent auto-close pure informational tickets or send a templated acknowledgment, while anything touching a refund, a churn risk, or a legal question still escalates to a person.

Follow-ups are where revenue quietly leaks, and they suit an AI workforce well because the discipline of touching a lead five or six times is the first thing a busy team drops. The agent can watch the pipeline, spot the threads that went silent past their reply window, and draft the next contextual touch, pausing the sequence when a prospect replies so no one gets an awkward automated nudge after they already answered. Because these messages go out under your name, keep the send human — this is comms, and comms keeps a person — but you can mandate the drafting and the queuing so every good lead always has a next touch waiting for approval.

Reporting and reconciliation round out the map, and they split along the same line. Reporting is mostly assembly: pulling the weekly numbers, call notes, and pipeline status into one prepared packet is exactly the kind of reversible, well-defined work that can graduate to a standing mandate quickly, since the cost of a wrong number in a draft report is a correction, not an incident. Reconciliation is the opposite discipline. Let the AI draft the matches, propose categories for predictable lines, and flag the anomalies and duplicates a human would miss across thousands of transactions — but the posting to the ledger stays behind a person, permanently. The pattern that repeats across all four functions is the same one this whole reset is built on: hand over the reading, drafting, and flagging; keep the decision, the send, and the money.