The Wrong Automation Conversation: Why Medtech Keeps Solving the Wrong Problem

The automation conversation in medical device distribution is stuck on a question that sounds right but leads nowhere useful. Every conference panel, every vendor pitch, every internal strategy session opens the same way: “What can we automate?”

The better question — the one that actually restructures costs instead of shifting them — is: “What work exists only because two systems don’t share context?”

That distinction is the difference between bolting incremental tools onto a broken architecture and eliminating the architecture’s need for human translation. It’s also the difference between companies that scale and companies that hire.

The bolt-on fallacy

The industry’s current approach to automation follows a pattern anyone who has watched enterprise tech adoption in other sectors will recognize. Somebody identifies a problem: manually re-typing order data from emails into an ERP. Somebody buys a point solution — OCR, an RPA script, a chatbot for customer service. The solution addresses the symptom. The underlying cause, that field communication and systems of record exist in incompatible formats, stays untouched.

BCG’s Medtech Companies Must Move Faster on GenAI (2023) makes the structural case directly: GenAI is “uniquely suited to tasks that incorporate a high volume of unstructured data inputs and generate coherent output in the form of text, code, images, and video,” and it can “relieve the burden of tasks that are highly repetitive but not fully automated because of workflow constraints” (BCG, 2023, p. 3). That is medtech back-office work described line by line — PDF purchase orders, photographed charge sheets, text-message inventory updates, the work that sits between unstructured field communication and structured ERP data. The leverage is in the unstructured layer, not in better OCR for structured inputs.

The Cognizant/Microsoft survey of 200 medtech decision-makers makes the gap even clearer. Ninety-one percent are enthusiastic about AI, but adoption is “most advanced in R&D and manufacturing, least mature in commercial operations” (Cognizant/Microsoft, 2024). The highest-volume, most repetitive operational process in distribution — order-to-cash — remains largely unaddressed.

The reason is structural. McKinsey describes the medtech operating model as decades of “siloed, functional organizations with complex matrices and diffused decision making,” where efficiencies are “deeply interconnected across functions, requiring an evolved operating model and cross-functional collaboration to unlock” (McKinsey, The Transformation Imperative in Medtech, 2025). Translation: the problem isn’t that nobody has built a good OCR tool. The work exists because the architecture was never designed for information to flow between the field and the system without a person in the middle.

RPA automates the translation. The right approach removes the need for translation.

Lessons from the headcount experiments

The technology sector has been running a large-scale natural experiment on what happens when companies replace coordination headcount with systems that share context natively. The results don’t match the headlines.

Shopify cut 20% of its global workforce in 2023, its second significant reduction after eliminating 10% the year before. Subsequent revenue growth appeared to validate the thesis: automation could absorb the coordination work human staff had been performing.

Klarna took the thesis further, replacing roughly 700 customer service agents with AI tools built on OpenAI’s models. The CEO publicly celebrated the efficiency gains. Then customer service ratings started declining. Complaints went up. By early 2025, Klarna was rehiring human agents to handle interactions the AI couldn’t manage. The CEO acknowledged that “cost unfortunately seems to have been a too predominant evaluation factor.”

Block — formerly Square — cut nearly half its workforce. The pattern repeated across dozens of tech companies.

The lesson isn’t that automation fails. It’s that automation only works when it targets work that requires no judgment, and the boundary between judgment-required and judgment-free work is harder to identify than most executives assume. BCG’s research makes the point directly: “The highest-ROI implementations augment frontline worker capability rather than replace headcount” (BCG, 2023).

That’s the rule most leaders get wrong on the first pass. They try to remove the people instead of removing the work those people shouldn’t have been doing.

What the transaction data actually shows

This is where aggregate tech narratives meet the specific reality of medical device distribution.

Deviceflow has processed over $3 million in medical device orders over the past two months, growing double-digits month over month. The transaction data shows a consistent distribution:

Sixty to seventy percent of inbound orders follow one of five predictable patterns. Same format, same fields, same sequence. A person processes these identically every time because there’s exactly one correct way to handle them. These orders shouldn’t require human intervention at all.

Twenty to twenty-five percent are variations on those base patterns: a missing field, an unusual quantity, a slightly different format. These need a judgment call, but a lightweight one. Flag the exception, fill the gap, confirm with the rep, process.

Five to ten percent are complex cases: custom pricing, split shipments, unusual product configurations, new account setups. These are the transactions where experienced ops people earn their salary. They need context, relationship knowledge, and judgment no system should try to replace.

The problem, consistently, is that most distributors have their best people spending the majority of their time on the first category. Work that requires zero judgment absorbs the most expensive labor. Not because anyone designed it that way, but because nobody designed it at all. The systems evolved. The work accumulated. Headcount grew to match.

The root cause sits in the format mismatch itself. Field reps send charge sheets as photos. Hospitals email PDF purchase orders. Inventory moves by text message. The information exists — it’s just trapped in formats no system can read without a human intermediary. As one Fortune 500 ortho senior director put it on a recent call: “An ERP is a financial system.” It can tell you a PO is unmatched. It can’t read the PDF, find the case, check the contract price, and chase the missing lot number.

For a deeper look at what that translation work looks like at the case level, see The Charge Sheet Problem and The Visibility Gap.

Medical device distribution is different — and that matters

The tech companies that cut headcount aggressively operated in a specific context: high-margin software businesses where a dropped customer interaction is an inconvenience, not a clinical event. Medical device distribution runs under different constraints.

When a $50,000 surgical case depends on a text chain between a rep, a coordinator, and an ops team, the stakes of changing the process feel existential. A dropped order doesn’t mean a customer churns to a competitor’s subscription. It means a surgeon walks into an OR without instruments. A patient’s procedure gets delayed. A hospital relationship built over years takes the hit.

That’s why the industry’s risk tolerance for operational change approaches zero. It’s also why the “automate everything” thesis imported from Silicon Valley is precisely wrong for this market.

The right approach is surgical, not sweeping. Automate the 60-70% of transactions that follow predictable patterns and require zero judgment. Flag the 20-25% that need lightweight human review. Leave the complex 5-10% in the hands of experienced staff, and free those staff from the repetitive work that currently prevents them from applying their expertise where it matters.

Salesforce’s 2026 State of Sales report (cross-industry, not medtech-specific) found reps spend 60% of their time on non-selling work, with 11% of the workweek going to manually entering data (Salesforce, 2026, p. 8). In medical device, that work doesn’t sit cleanly between meetings — it lands after each case, on top of the four manufacturer lines a top rep already carries in their head. Take the back-office burden off the rep and you’re not just running a leaner ops shop; you’re returning hours of pure selling time to the part of the team that earns the next case. That’s not an efficiency story. That’s a revenue story hiding inside an operations problem.

This isn’t the automation conversation happening at industry conferences, where panels still debate barcode scanning versus RFID and “digital transformation” means buying a new ERP. The actual leverage point is upstream, in the space between communication and system, where information changes format and context gets lost.

What’s safe to automate today

The Klarna reversal is instructive. The failure wasn’t automating customer service. It was automating customer service without sorting which interactions needed judgment and which didn’t. Every distributor can do that sort before buying anything.

Walk through your back office and put each type of work in one of three buckets.

Safe to automate now — work that requires zero judgment:

• Order intake from your top 10 accounts, where the format is predictable
• PO-to-contract-price matching against a known price book
• Inventory location updates from rep texts (“restocked 3 of SKU X at [hospital]”)
• Case confirmation emails back to reps
• Invoice generation for bill-only POs against existing contracts
• ERP entry from PDFs that follow known templates

Automate with human review — work that’s mostly pattern but occasionally novel:

• Orders from new accounts where the format hasn’t been seen before
• Exception cases with missing fields or unusual quantities
• Split shipments
• Requests for custom pricing outside existing contracts

Keep human — work that requires relationship knowledge:

• New product launches and contract setup
• Distributor relationship changes
• Hospital committee approvals
• New surgeon onboarding or complex first-case setups

Most distributors haven’t run this exercise. When they do, the first surprise is usually how much bucket-one work is sitting in bucket-three hands.

Where to start

The starter path that works without buying anything, for a distributor reading this on a Monday morning:

Weeks one and two — audit. Have the ops team log what they do in 15-minute increments for ten business days. Tag each entry as pattern-matching data entry, exception handling, or complex judgment. Your own 60-20-10 split will surface within two weeks. It almost always lines up with the industry pattern, but the specific categories that dominate vary by distributor. If you want a structured version of this exercise, The Visibility Gap walks through it at the case level, and The PO Mess does the same for the billing side.

Weeks three and four — pick one pattern. Identify the single highest-volume, most standardized order type flowing through your back office. For most distributors it’s one of three things: bill-only POs from the top 10 hospitals, stocking orders that replenish against known PAR levels, or charge sheets from a handful of surgeons with predictable case mixes. Don’t try to automate three patterns at once. Pick one.

Weeks five through eight — automate that one pattern end-to-end. Order intake, PO-to-contract matching, ERP entry, invoice generation. If you’re building internally, this is a two-developer sprint. If you’re buying, it’s a single-vendor evaluation on one workflow, not a platform decision. Keep the scope narrow enough that success or failure is visible in four weeks, not four quarters.

Weeks nine through twelve — measure. Three numbers matter. Case-to-invoice cycle time should compress by 50% or more. Ops hours per order in that category should drop by 75%. Error rate should match or beat the manual baseline. If any of those miss, the pattern wasn’t as standardized as you thought, and the human-review window needs to widen before expanding.

After 90 days — expand. Pick the next pattern. By the end of the second quarter, three patterns should be automated and the ops team should be reviewing exceptions rather than processing volume. That’s when operational leverage starts compounding.

For most distributors, the second pattern after order intake is field inventory — trunk stock, consignment, and the endless count reconciliation work that sits between rep texts and the ERP. The playbook is the same: audit, pick one pattern, automate end-to-end, measure.

The companies that will navigate this correctly aren’t the ones asking “what can we automate?” They’re the ones asking “where does context get lost between the field and the system?” and building connective infrastructure to make sure it doesn’t.

The industry will move slower than tech. Given the stakes, it should. But the direction isn’t in question. The only variable is whether individual companies design the transition deliberately or wait until margin compression writes the memo for them.