# Back-office automation: the business that updates itself > A self-updating business is not autonomous agents but a back office that keeps its own data in order. The JD example and what a Ukrainian SMB can take from it. Source: https://auspex.company/en/blog/self-updating-business/ [Blog](/en/blog/) / Field Notes # The business that updates itself: scaling without hiring doesn't start with agents By: Andrew Maryasov · Aug 11, 2026 · 8 min read A self-updating business is a company whose back-office functions, such as product, inventory and customer-base management, maintain themselves with the help of AI, so the volume of operations grows faster than headcount. Jack Clark coined the term in his newsletter, Import AI #464. It was misread almost at once: people hear “AI does the work by itself,” when the point is something else — a data layer that stops decaying. That difference decides whether back-office automation will work for you. ![Three levels of a self-updating business drawn as steps: normalized reference lists, checks at entry, AI-proposed changes with human confirmation](/images/blog/self-updating-business-en.webp) Scaling without hiring doesn't start with agents ## What is a self-updating business? Clark described the phenomenon while commenting on an engineering paper from the Chinese retailer JD: modern AI tools make it possible to build companies that have intelligence woven into their back-office functions — inventory management, for example. Such companies operate at far larger scales than was possible before, and they can also update themselves and keep learning, often without large amounts of human oversight. > Modern AI tools let us create businesses that have intelligence woven into their back-office functions, like inventory management. > > — Jack Clark, Import AI #464 The key word here is “woven.” Not “placed on top,” not “plugged in on the side.” What most companies call automation lives in a separate layer: a robot collects requests once an hour, a bot answers a routine question. Take that layer away and the business keeps running, just more slowly. A self-updating system is built the other way around: it is part of how the company knows what is going on inside it at all. ## What it looks like at JD: why does the ontology live outside the model? JD is often called the Amazon of China. According to the figures Import AI cites, the company has 700 million users, millions of merchants and a catalog of tens of billions of SKUs. Managing a catalog like that by hand is impossible in principle — it's a matter of arithmetic, not diligence. JD described the system that does this job in a research paper: the Oxygen AI Item Center. According to the company, it covers tens of thousands of categories and processes hundreds of millions of item updates per day. Those numbers have nothing to do with a 20-person business. What's worth borrowing is not the scale but the architectural decision behind it. JD moved the ontology — the structure of knowledge about what a product is, which category it belongs to and which attributes it has — out of the model and into a separate knowledge base. That lets the ontology be updated continuously without retraining the model. The model itself was left with a narrow role: the system first searches the knowledge base for relevant entries, and the model only determines whether the item matches the entry it found. The paper says plainly why it was done this way: it substantially reduces task complexity and lowers the risk of hallucinations. AI here does not “run operations” in place of people. It keeps up to date the description of the business that every other process relies on. What gets updated is the system's knowledge of itself, not its behavior. ## Why isn't this about “an agent that will do everything for you”? The same issue of Import AI has numbers worth keeping at hand when someone is selling you an end-to-end autonomous agent. On OSWorld 2.0, a benchmark where AI systems carry out long real-world tasks on a computer, the strongest configuration at the time of measurement (June 2026) scored 20.6% on fully correct completions and 54.8% on partial ones. On the benchmark's leaderboard as of October 2026, Claude Opus 5 reaches 44.33% on the updated task set — still fewer than half. The authors note that performance drops sharply as tasks grow longer, and that agents struggle most with exactly what operational work is made of: recovering hidden state, keeping track of many items, resolving conflicting information and adapting to requirements that have changed. The second dimension of the same picture is the Remote Labor Index, a joint benchmark from the Center for AI Safety and Scale Labs (Scale AI). Here models carry out real paid freelance projects in full, from start to finish, and the work is accepted by a human client, not an automated test. At launch in October 2025 the best result was 2.5% of projects accepted; in July 2026 it was 15.8%. That is more than a sixfold increase in under a year, and the direction is unmistakable. But even as of October 2026 the best result is about 21%. A bet on an agent that runs an end-to-end process on its own rests, as of October 2026, on exactly these two measurements: about 21% of projects accepted on the Remote Labor Index and 44.33% fully correct completions on the updated version of OSWorld 2.0 — both below half. A bet on a self-updating data layer is a bet on a task that JD narrowed down to “match or no match” for a reason: narrow, repeatable and cheap to verify. Automation takes root where checking the result costs next to nothing. ## What actually limits growth without hiring? When an SMB hits a ceiling, the cause gets described as a shortage of hands. Look closely at specific processes and you almost always see a different picture: what's missing is not hands to do things but an up-to-date description of the business. A large part of that description isn't written down anywhere — it lives in employees' heads. - Reference lists of sources, statuses and categories where half the values appeared by accident and no longer mean anything. - The same customer in several records, because they came in through the website, through Instagram and by referral. - A product catalog where names and specifications are filled in by whoever got there first, however they understood them. - Pricing exceptions and arrangements that one particular sales rep remembers and nobody else does. - Deal loss reasons picked at random, because the field is required and the options don't describe reality. Volume grows, and with it grows the number of unwritten exceptions that someone has to keep in their head. Hiring becomes unavoidable not because more customers require more hands, but because every new knowledge holder is needed to keep the system understandable. You are hiring people to serve as living documentation. A self-updating back office breaks exactly this link. If reference lists normalize themselves and duplicates never make it into a report, the description of the business stops depending on who is in the office today. Volume then grows faster than headcount, because the work that used to require hiring is gone. ## Level 1: normalized reference lists The most boring level and the most profitable one. Single, agreed values for lead sources, customer and deal statuses, products, regions, loss reasons and request types. Separating the legal form from the brand name is worth doing as its own step: without that, “TOV Romashka,” “Romashka” and “Romashka LLC” count as three different customers, and every repeat-sales report lies. There is no AI here yet — but without this level the rest doesn't work. ## Level 2: checks at entry A database goes bad at the moment a record is created, not gradually. So the checking moves to the point of entry: deduplication by phone and email before a second record appears, immediate normalization of number and name formats, auto-filling of required fields from the communication history. It's the same idea as at JD, only at SMB scale: the system compares a new object against the existing reference list and settles a narrow question about a match. ## Level 3: AI-proposed changes with human confirmation The highest level available to an SMB today without excessive risk. AI scans the database regularly and proposes changes: merge these records, add a value to the reference list because 40 deals in a quarter used “other,” mark a contact as invalid. The decision stays with a person, but that person works with a list of proposals, not with a database of ten thousand records. This is where the self-updating property appears — and where it is safe, because each proposal is checked in seconds. ## Where to start: what sequence does Auspex follow? - Data audit: the share of duplicates, empty required fields, inactive contacts and inconsistent names. - Data rules: required fields, phone and email formats, canonical company spellings, a closed list of statuses. - Cleanup and deduplication: merge existing duplicates, archive what's no longer relevant, fill in records from the history. - Tidying up reference lists: single values for sources, categories, regions, products and loss reasons. - Automating routine work: lead distribution, reminders, follow-ups, reports, monitoring of stalled deals. - Regular checks: a weekly review of new records and a full scan of the database every month. Order matters more than speed here. Steps 1–4 don't make for an impressive demo, and they are the ones you'll most often be advised to skip in order to “get to AI faster.” Skipped, they turn any later automation into an accelerator of disorder: the lead distribution rule works flawlessly and distributes duplicates, the report builds instantly and counts one customer as three. ## Test: is your back office already self-updating? - How long will it take to prepare a quarterly repeat-sales report — and will the CFO trust it without a manual check? - What happens to data quality if your most experienced sales rep goes on leave for three weeks? - How many days does a new employee need to understand which status to set in a non-obvious situation? - Who will notice that a duplicate customer has appeared — the system, or an annoyed customer who got called twice? - When did anyone last review the list of values in the loss reasons reference list? If the answers to these questions come down to specific people rather than specific mechanisms, the back office runs on the team's memory. That works up to a certain volume; after that, every additional percent of growth has to be bought with hiring. ## What does this give you in numbers — and what don't we promise? There is no universal multiplier here, and anyone who names one before an audit is making it up. The effect depends on how neglected the database is at the start and how many operations pass through it every day. What you can actually measure before and after: the share of duplicates in the database, the hours spent preparing regular reports, the time it takes a new sales rep to start working independently, the share of deals with fully completed data. These metrics can be captured in a week and don't have to be taken on faith. A telling example of a cumulative effect is GRO, a Ukrainian retailer and service center for Apple devices (the company agreed to publication). Auspex has supported the company as its CRM integrator since 2021: system implementation, sales automation, CRM-to-ERP synchronization. Over that period GRO grew from roughly 10 orders a month to 10 a day. That is the result of several years of consistent work on processes and data, not the outcome of a single implementation. The realistic horizon for changes like these is measured in years, not quarters. ## What's the next step? Auspex is a CRM implementation and business process automation company for SMBs. We start not with choosing a tool but with a question: exactly which data your business loses every day, and how many human hours go into keeping the system in a sane state. Once that layer is sorted out, AI agents built on top of it for support or lead qualification really do have something to stand on — that line of work belongs to Grow2.ai, our AI division. Start with an audit: 30 minutes of conversation, and you'll see which operations in your back office can maintain themselves and where growth will keep costing you new hires. Read next - [Grow2.ai — our AI division: agents for your process](https://grow2.ai/en/) - [Where to start automating in a small business: 5 boring processes](/en/blog/where-to-start-automation-smb/) - [When CRM automation really pays off: three stories](/en/blog/crm-automation-payback/) - [GRO: from 10 orders a month to 10 a day](/en/blog/gro-ecommerce-crm-erp-case/) ## Sources - [Import AI #464 — Jack Clark on “self-updating businesses” (a look at the JD paper)](https://importai.substack.com/p/import-ai-464-fables-writes-gpu-kernels) - [Remote Labor Index — Center for AI Safety and Scale Labs](https://www.remotelabor.ai/) Andrew Maryasov Founder, Auspex Founder of Auspex. Implementing CRM and automating business processes since 2015 — 1,500+ projects for 1,200+ companies. [amaryasov.expert →](https://amaryasov.expert) ## Frequently asked What is a self-updating business? It is a company whose back-office functions — product, inventory and customer-base management — keep themselves current with the help of AI, without constant manual intervention. Jack Clark coined the term in his newsletter, Import AI #464. The practical sign is this: the volume of operations grows faster than the number of people on staff, because the system itself keeps in order the description of what is going on inside it. How is this different from ordinary automation? Ordinary automation speeds up actions: it sends reminders, builds a report, distributes incoming requests. Self-updating works at the level of data: the system keeps reference lists current, removes duplicates and fills in records. The difference is critical, because automating actions on top of bad data simply multiplies errors faster. Does this require autonomous AI agents? No, and today they are not worth betting on for end-to-end operations. As of October 2026, on the OSWorld 2.0 benchmark the strongest configuration produces 44.33% fully correct completions of long real-world tasks, and on the Remote Labor Index the success rate for end-to-end projects is about 21%. The approach that works is narrow, verifiable tasks: compare a record against a reference list, find a duplicate, propose a change for a person to confirm. Where should a small company start with back-office automation? With a data audit: count the share of duplicates, empty required fields and inconsistent company names. It takes a few days and gives you the baseline against which every later effect is measured. Next come data rules, cleanup and normalization of reference lists, and only then automation of routine operations. The order is essential: automation launched on an uncleaned database dutifully sends reminders for duplicate records and builds reports nobody can trust. How much does it cost and how fast does it pay back? There is no universal figure — the effect depends on the state of the database and the volume of daily operations, so any multiplier named before an audit is made up. The work is split into stages: audit and data rules, then cleanup and reference lists, then automation. Each stage has its own measurable result, so the decision about the next step is made on the evidence of the previous one, not on a promise given at the start. Does this really let you scale without hiring? It removes one specific reason to hire — the need for people who keep unwritten rules and exceptions in their heads. Hiring for new skills, sales or production doesn't go away. The point is that growth in the volume of operations stops automatically meaning growth in administrative headcount. ## Business automation digest 2–3 emails a month — what actually works in CRM and automation. No spam. Unsubscribe in one click. [Free CRM audit →](/en/contact/) ↳ FREE AUDIT 30 min · free · ↓ ## Ready to Get Your Sales Under Control? Free process audit — 30 minutes, and you'll receive: 1,200+ companies since 2015 · 30-day median go-live 01 A map of where you're losing leads 02 Priorities: what to automate first 03 An implementation plan for 2-4 weeks We'll call within 24 hours. The plan is yours, even if we don't work together.