1
WHITEPAPER
Using AI to Optimize
Lab Operations: Begin with
the Bottlenecks
Most labs aren’t missing the AI tools. They’re missing the starting point. This guide shows
how to use AI where your lab needs it
AI tools have made early inroads into lab work. Scientists use chatbots and language models to draft reports, summarize papers, and write custom code. But
core operations still rely on fragmented data and manual processes. Even when AI models are available, teams often struggle to apply them where they could have the most impact.
That disconnect drew a full house to the “AI in Action”
workshop at the 2025 Lab Manager Leadership Summit.
Led by Yahara Software, a Madison-based firm specializing in custom software for scientific and clinical labs, the
session guided attendees in building AI plans personalized
to their lab needs.
“AI is not a solution to every problem,” said Garrett Peterson, Yahara’s chief commercial officer. “Start by defining
the task, understanding the return on investment, and
choosing the tool that fits.”
This guide distills practical insights from the session to
help labs focus their AI efforts where they matter most.
Where AI Is Already at Work
Early in the session, the Yahara team made a key distinction: the most measurable gains in labs are coming from
machine learning, not large language models.
James Smagala, bioinformatics manager at Yahara,
described a project where a machine learning model now
classifies particle images and flags anomalies, replacing a
tedious manual step. He emphasized the value for visual
data: “Think of analyzing, segmenting, and classifying
cells. Wouldn’t it be better if artifacts or inconsistencies
jumped out automatically?”
Smagala pointed to quality control and compliance as
strong starting points for automated solutions. These areas
2
involve high-volume, repeatable tasks with well-defined
outcomes, making them especially well suited to machine learning.
Machine Learning in the Lab
Adam Steinert, chief technical officer at Yahara, explained that machine learning is valuable when problems
become too complex for rule-based logic. In cases such as
image classification, writing fixed instructions to identify
every variation quickly falls apart.
“You might hardcode features to recognize a dog—ears,
snout, tail—but change the breed or the camera angle, and
those rules stop working,” he said. Machine learning takes
a different approach, learning from large sets of examples
rather than relying on predefined rules.
These models operate on probabilities, not absolutes.
“They are not always perfect or predictable,” Steinert
noted, “but they can solve problems traditional algorithms
cannot.” The key is exposure. As models see more data and
receive more feedback, they improve significantly.
Labs apply machine learning through four main
approaches:
• Supervised learning trains models on labeled examples.
It is used for tasks such as classifying images, predicting
lab values, and detecting sample errors.
• Unsupervised learning identifies hidden patterns in
unlabeled data, helping researchers cluster unknown
samples or reveal unexpected relationships.
• Reinforcement learning teaches models by rewarding successful actions, often used in robotics, dynamic
scheduling, and lab automation.
Deep learning relies on layered neural networks to recognize complex patterns. These models power applications
such as image analysis, speech recognition, and large
language tools. They require large datasets and significant computing resources but drive many of today’s most
advanced AI systems.
Each method addresses a different type of complexity.
Steinert emphasized that machine learning succeeds not
by finding perfect answers, but by building systems that
improve through iteration and exposure to real-world data.
Frame the Problem Like a Scientist
Peterson told attendees to approach AI projects like
experiments: set a clear objective, control the inputs, and
measure the results.
To surface those pain points, teams used ‘lean’ principles
to identify inefficiencies in their lab workflows. They listed
common forms of waste—overproduction, waiting, excess
motion, underused staff—and wrote down where those
showed up in their own labs. Each problem got a quick cost
estimate to size its impact.
Once a candidate task surfaced, the next step was to
define it precisely: What triggers it? What should happen? Where does it fail? Smagala noted that framing by
function—not by technology—leads to stronger pilots.
He outlined three functions that align well with machine-learning tools:
• Classification: sorting or labeling inputs, such as
separating true peaks from artifacts or routing samples
by metadata.
• Transformation: converting data formats, such as turning scanned forms into structured records or cleaning
log files.
• Generation: producing structured drafts from templates, such as summarizing SOPs or creating onboarding guides.
Precise scoping helps teams predict where AI might
succeed or fail under lab conditions.
Planning and Launching an AI Pilot
Peterson laid out three realities every lab should accept
before sketching an AI pilot:
1. Data comes first.
“AI needs volume,” he said. The bottlenecks worth automating are the ones backed by large, reliable datasets, or a
clear plan to gather them.
2. You probably won’t build the model.
“Even when we deploy AI for instrument vendors, we
almost never train from scratch,” Peterson noted. Off-theshelf models, fine-tuned on outside data, deliver results
sooner and cost less. Integration, not invention, moves
pilots forward.
3. AI is a tool, not the solution.
Peterson warned against chasing problems just because
they “sound like AI.” Start with the business case, confirm
the payoff, then pick the right tools in the right order. Often that means fixing basic data hygiene and workflow gaps
before any algorithm can add value.
With the groundwork set, the Yahara team recommends launching a focused two-week pilot tied to a single
metric—such as cutting error rates, trimming turnaround
time, or reducing manual review.
3
To learn more, visit: yaharasoftware.com/labmanager
Preparing Lab Data for Machine Learning Models
Once a pilot is in motion, the next step is to get hands-on
with the data. “These steps don’t always happen in order,”
said Steinert. “But data is foundational. You need to ask early:
Do I have it? Can I get to it? And is it even the right kind?”
He stressed that for process optimization, the missing
piece is often infrastructure—labs may not track how long
tasks take, how instruments are used, or how staff move
between steps. “Sometimes the real work is setting up systems that can collect that in the first place,” he said.
Once data is in hand, exploration begins. Teams break it
down, visualize it, and look for usable patterns. “You might
not know the key features at the start,” said Steinert. “You
find them by working with the data—and sometimes by
realizing what’s missing.”
From there, cleanup and preparation take over: correcting errors, formatting fields, balancing inputs. This phase
takes time and expertise, but ensures the model can handle
real lab data.
“The key takeaway here? Iterate,” said Steinert. “One of
the most common ways these projects fail is by expecting
them to work on the first try.” Smagala added that even production models need tuning as lab conditions evolve. Treat
AI like any other scientific method: test, adjust, improve
Success Factors for AI in Scientific Labs
AI work is, at its core, informatics. “Everything you’ve
learned from past IT rollouts? It all applies here,” Peterson
told the group. That includes infrastructure, integration,
and buy-in from IT. For labs without that experience, his
advice was simple: “Find people with the scars and the
stories. They’ll help you avoid the traps.”
Internal expertise is just as critical. Steinert and Smagala
underscored the need for domain knowledge during data
preparation, labeling, and validation. “You need people
who understand the edge cases,” Smagala noted. Without
them, models learn the wrong signals—and in regulated
settings, that means failed audits. “If the process can’t be
verified, it won’t hold up,” Peterson warned.
Collaboration can unlock more than internal effort
alone. Peterson pointed to vendors holding high-value
data, such as usage trends, inventory logs, and workflow
patterns. “If that information moves beyond individual
labs, the whole system gets smarter,” he explained. Teams
should assess where external partnerships could provide
missing data, tools, or expertise.
AI projects succeed when they’re built with structure,
focus, and the right support. Yahara helps labs turn pilots
into reliable, validated systems. Learn more at yaharasoftware.com/labmanager.
Meet the Experts
Garrett Peterson
Chief Commercial Officer, Yahara Software
With over 35 years of experience, Garrett
helps labs improve their processes and
adopt the right digital tools to meet today’s
challenges. He leads Yahara’s efforts to
build software that supports both science
and business needs.
James Smagala
Bioinformatics Manager, Yahara Software
James specializes in biochemical
assays, NGS data management, and
bioinformatics. He bridges wet-lab
workflows with digital tools, designing
global platforms that streamline data
capture, analysis, and collaboration.
Adam Steinert
Chief Technical Officer, Yahara Software
Adam leads Yahara’s technical strategy,
drawing on deep experience in scientific
software, instrumentation, and applied
machine learning. He spearheads efforts
to bring practical AI tools into labs and
accelerate R&D innovation.
AI Pilot Planning Worksheet
A companion resource for lab managers planning
practical AI applications
Section 1 - Identify the Task
What lab task or workflow are you targeting?
.............................................................................................................................................................................................................................................................................................................................................................
Why is it worth improving? (e.g., time lost, error rate, backlog)
.............................................................................................................................................................................................................................................................................................................................................................
Time per run (hrs): .......................................................................................................... Error rate (%): .......................................................................................................................
Other impact: ....................................................................................................................................................................................................................
What triggers this task? ..............................................................................................................................................................................................
What would a successful outcome look like? (Be specific and measurable)
.......................................................................................................................................................................................................................................................
.......................................................................................................................................................................................................................................................
Section 2 - Define the Function
Which of the following best matches how this task works? Choose one.
Classification (flagging, routing, detecting)
Transformation (formatting, extracting, converting)
Generation (drafting, simulating, pre-filling)
Briefly explain why this fits:
.......................................................................................................................................................................................................................................................
.......................................................................................................................................................................................................................................................
Section 3 - Assess the Data
List any relevant data sources: .............................................................................................................................................................................
Source: ................................. Format: ....................................... Volume: ...................................... Accessible? Y N
Gaps or restrictions: ......................................................................................................................................................................................................
Do you need new systems or infrastructure to capture missing data? Y N
If yes, describe: .................................................................................................................................................................................................................
Section 4 - Select the Approach
Which tool or method might fit this task?......................................................................................................................................................
OCR or vision model Off-the-shelf classifier Language model or chatbot
Other: .................................................................................................................................................................................................................................
Why do you think this approach fits the problem?
.......................................................................................................................................................................................................................................................
.......................................................................................................................................................................................................................................................
Section 5 - Plan the Pilot
Start Date: .................................................................................................................................... End Date: .............................................................................................
Success Metric: ................................................................................................................................................................................................................
Baseline: ................................................................................................. Target: .................................................................................................
Technical lead: .................................................................................................................................................................................................................
Domain expert: .................................................................................................................................................................................................................
Where will results show up in the workflow?
.......................................................................................................................................................................................................................................................
.......................................................................................................................................................................................................................................................
Section 6 - Traceability and Support
Owner for long-term maintenance: .................................................................................................................................................................
Data storage location: ................................................................................................................................................................................................
Access control in place? Y N.......................................... Version control system in place? Y N
Are there SOPs or regulatory guidelines that apply? ............................................................................................................................
Section 7 - Validation Plan
Test set held out? Y N
Performance will be tracked using: ...................................................................................................................................................................
Known sources of bias (data, labels, usage): ............................................................................................................................................
Section 8 - Estimate ROI
Hours saved/month Baseline: ..................................... Target: ......................................... Actual: ...............................
Error rate (%) Baseline: ..................................... Target: ......................................... Actual: ..............................
Cost impact ($) Baseline: ..................................... Target: ......................................... Actual: ..............................
Section 9 - Reflect and Decide
What did the pilot show?............................................................................................................................................................................................
Next step: ..............................................................................................................................................................................................................................
Scale and integrate Refine and rerun Archive and revisit later
